CN110162698A - A kind of user's representation data processing method, device and storage medium - Google Patents
A kind of user's representation data processing method, device and storage medium Download PDFInfo
- Publication number
- CN110162698A CN110162698A CN201910312225.3A CN201910312225A CN110162698A CN 110162698 A CN110162698 A CN 110162698A CN 201910312225 A CN201910312225 A CN 201910312225A CN 110162698 A CN110162698 A CN 110162698A
- Authority
- CN
- China
- Prior art keywords
- feature
- target
- user
- data information
- moment
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The embodiment of the present invention discloses a kind of user's representation data processing method, device and storage medium, wherein method includes: that response for the first of target application pulls request, obtains first and pulls the corresponding feature update duration of request;It includes multiple first moment that feature, which updates duration,;Target network model is updated obtained hiding feature in duration in a upper feature to be determined as initially hiding feature;Corresponding business datum feature of multiple first moment is obtained from business data characteristics vector, and corresponding user's Figure Characteristics of multiple first moment are obtained from user's Figure Characteristics vector;Feature, corresponding business datum feature of each first moment and user's Figure Characteristics are hidden according to initial, determine that target network model hides feature in the target that feature updates in duration;The target hides feature and pulls the user for requesting corresponding target user portrait information with described first for characterizing.Using the present invention, it can be ensured that the accuracy and completeness of user's representation data.
Description
Technical field
The present invention relates to Internet technical field more particularly to a kind of user's representation data processing methods, device and storage
Medium.
Background technique
Existing recommender system is recommending service data information (for example, the internets such as article, song are believed to target user
Breath) when, customer attribute information when can be registered based on the target user carries out data recommendation.It is understood that existing
Recommender system the target user can be registered the target application when institute typing relevant information (for example, gender: male, the age:
18, label: swordsman's class) it is referred to as the customer attribute information of the target user, it can relevant information when by the registration determines
For user's portrait information of the target user.However, over time, the user of the target user draws a portrait information may
It changes, for example, the hobby of user can change, but the target user is still registered Shi Suolu by the recommender system
The relevant information entered is considered as user's portrait information of the target user, believes so that it cannot ensure that the user of the target user draws a portrait
The accuracy and completeness of breath.
Summary of the invention
The embodiment of the present invention provides a kind of user's representation data processing method, device and storage medium, it can be ensured that user
The accuracy and completeness of representation data.
On the one hand the embodiment of the present invention provides a kind of user's representation data processing method, which comprises
Response pulls request for the first of target application, when acquisition described first pulls request corresponding feature update
It is long;It includes multiple first moment in duration that the feature, which updates,;
Target network model is updated obtained hiding feature in duration in a upper feature to be determined as initially hiding feature;
Obtain corresponding business datum feature of the multiple first moment from business data characteristics vector, and from
Corresponding user's Figure Characteristics of the multiple first moment are obtained in the Figure Characteristics vector of family;
According to the initially hiding feature, corresponding business datum feature of the multiple first moment and described
Multiple first moment corresponding user's Figure Characteristics determine that the target network model updates in duration in the feature
Target hides feature;The target hides feature and pulls the user for requesting corresponding target user with described first for characterizing and draw
As information.
Wherein, it is to pull the corresponding first request time stamp of request by described first to draw with second that the feature, which updates duration,
It takes determined by the corresponding second request time stamp of request;Described second pulls the upper drawing that request pulls request for described first
Request is taken, described second pulls request for obtaining the multiple history recommending data information for having incidence relation with target user;
The method also includes:
It updates in duration, will be deposited between the target user in the multiple history recommending data information in the feature
It is determined as target histories recommending data information in the history recommending data information of data-triggered relationship;
Based on each target histories recommending data information and the history service data information, the target network is constructed
The corresponding business datum feature vector of model;It include corresponding industry of multiple first moment in the business datum feature vector
Business data characteristics;The corresponding target histories recommending data information of one business datum feature;
Pass through each corresponding target user's attribute information of target histories recommending data information and the history industry
The corresponding historical user's attribute information of business data information, constructs the corresponding user's Figure Characteristics vector of the target network model;
It include corresponding user's Figure Characteristics of the multiple first moment in user's Figure Characteristics vector.
Wherein, described to be based on each target histories recommending data information and history service data information, construct the mesh
Mark the corresponding business datum feature vector of network model, comprising:
Each target histories recommending data information is added to the first business trigger list, and each target will be carried
First business trigger list of history recommending data information is as the second business trigger list;The first business trigger list and
Comprising having the history service data information of incidence relation with the target user in the second business trigger list;
In the second business trigger list, by the first moment corresponding target each of in feature update duration
History recommending data information is determined as target service data information;
The semantic feature of the target service data information is extracted, and using the semantic feature extracted as described in
The corresponding target service data characteristics of target service data information;
The corresponding historical data feature vector of the first business trigger list is obtained, and the target service data are special
Sign is spliced with the historical data feature vector, and spliced historical data feature vector is determined as the target network
The corresponding business datum feature vector of network model.
Wherein, it is described by the corresponding target user's attribute information of each target histories recommending data information and
The corresponding historical user's attribute information of the history service data information constructs the corresponding user's portrait of the target network model
Feature vector, comprising:
It is in the second business trigger list, the target histories recommending data information institute at each first moment is right
The customer attribute information answered is determined as target user's attribute information;
By the user property feature in target user's attribute information and the attention mould in the target network model
Type feature is combined, and the user property feature after combination is determined as the corresponding target of target user's attribute information and is used
Family Figure Characteristics;
Historical user's feature vector associated with the first business trigger list is obtained, and the target user is drawn
As feature and historical user's feature vector are spliced, and spliced historical user's feature vector is determined as the mesh
Mark the corresponding user's Figure Characteristics vector of network model.
Wherein, the multiple first moment includes historical juncture and object time, and the multiple first moment respectively corresponds
Business datum feature in include the historical juncture corresponding second business datum feature, the object time corresponding first
Business datum feature;The multiple first moment corresponding user's Figure Characteristics include the historical juncture corresponding second
The corresponding first user Figure Characteristics of user's Figure Characteristics, the object time;The historical juncture is the object time
Last moment;
It is described according to the initially hiding feature, corresponding business datum feature of the multiple first moment and
The multiple first moment corresponding user's Figure Characteristics determine that the target network model updates duration in the feature
Interior target hides feature, comprising:
By the target network model by the second user Figure Characteristics, the second business datum feature and institute
It states initially hiding feature to be merged, and using fused initially hiding feature as the target network model in the history
The history at moment hides feature;
By the target network model by the first user Figure Characteristics, the first business datum feature and institute
It states the hiding feature of history to be merged, and fused history is hidden into feature as the target network model in the target
The target at moment hides feature.
Wherein, the target network model includes the first resetting door, the second resetting door and updates door;
It is described by the target network model by the first user Figure Characteristics, the first business datum feature with
And the hiding feature of the history is merged, and fused history is hidden feature as the target network model described
The target of object time hides feature, comprising:
In the target network model, corresponding first weight parameter of the first business datum feature, first are obtained
Corresponding second weight parameter of user data feature, the history hide the corresponding third weight parameter of feature and the mesh
Mark the straggling parameter of network model;
By the first business datum feature, first weight parameter, the first user Figure Characteristics, described second
Weight parameter, the history hide feature, the third weight parameter and the straggling parameter as the target network mould
It is corresponding to export the corresponding gating parameter of the update door, the first resetting door for the corresponding input parameter of activation primitive in type
Gating parameter and the corresponding gating parameter of the second resetting door;
Feature and first business datum are hidden based on the corresponding gating parameter of the first resetting door, the history
Feature determines the corresponding candidate hiding feature of the first business datum feature in the target network model;
Feature and first user portrait are hidden based on the corresponding gating parameter of the second resetting door, the history
Feature determines the corresponding candidate hiding feature of the first user Figure Characteristics in the target network model;
According to the corresponding candidate hiding feature of the corresponding gating parameter of the update door, the first business datum feature,
The corresponding candidate hiding feature of the first user Figure Characteristics and the history hide feature, determine the target network
Model hides feature in the target of the object time.
Wherein, the method also includes:
The user that is characterized of feature information of drawing a portrait is hidden by the target, and multiple the are obtained from service database
One candidate data information;
Obtain the target user corresponding preliminary screening rule, and from the multiple first candidate data information sifting with
The first candidate data information that the preliminary screening rule is consistent is as the second candidate data information;
It selects to hide the use that feature is characterized with the target from all second candidate data information filtered out
The second candidate data information that family Figure Characteristics vector, the business datum feature vector are consistent is as data information to be recommended;
The target user is handed down to using selected data information to be recommended as target recommending data information.
Wherein, described to be handed down to the target use for selected data information to be recommended as target recommending data information
Family, comprising:
Each of selected data information to be recommended data letter to be recommended is predicted by the first clicking rate prediction model
Cease corresponding click probability;
The corresponding click probability of each data information to be recommended is ranked up, and from the data to be recommended after sequence
Target recommending data information is obtained in information, and gives the target recommending data information recommendation to the target user.
Wherein, described that the corresponding click probability of each data information to be recommended is ranked up, and after sequence
Target recommending data information is obtained in data information to be recommended, and the target recommending data information recommendation is used to the target
Family, comprising:
The click probability of each data information to be recommended is ranked up based on drop ordering rule, and after sequence
Screening clicks to be recommended data information of the probability greater than click probability threshold value as target recommending data in data information to be recommended
Information, and the target recommending data information filtered out is pushed to target user.
Wherein, the target recommending data information includes multiple sub- recommending data information;
The method also includes:
Described the is generated according to the click probability of every sub- recommending data information in the multiple sub- recommending data information
The corresponding predicted vector of one clicking rate prediction model;The predicted vector includes target prediction vector;The target prediction vector
The subsequent time of request is pulled in the multiple sub- recommending data information described first for describing the target user
The prediction probability that first sub- recommending data information is clicked;
Second in the multiple sub- recommending data information is directed in the subsequent time when the target user is recorded
When the click relationship of sub- recommending data information, the corresponding point of the first clicking rate prediction model is generated based on the click relationship
The amount of hitting to;The click vector includes the target point amount of hitting to;The target point amount of hitting to is for describing the target user in institute
State the actual probabilities that subsequent time clicks the second sub- recommending data information in the multiple sub- recommending data information;
The damage of the first clicking rate prediction model is determined according to the target point amount of hitting to described in the target prediction vector sum
Function is lost, and the loss function is optimized by gradient decreasing function, and the loss function after optimization is determined as mesh
Scalar functions, and the corresponding first clicking rate prediction model of the objective function is determined as the second clicking rate prediction model.
On the one hand the embodiment of the present invention provides a kind of user's representation data processing unit, described device includes:
Duration obtains module, pulls request for the first of target application for responding, and obtains described first and pulls request
Corresponding feature updates duration;It includes multiple first moment in duration that the feature, which updates,;
First obtains module, true for target network model to be updated obtained hiding feature in duration in a upper feature
It is set to initially hiding feature;At the time of the initial time is that a upper feature updates in duration;
Second obtains module, for obtaining corresponding industry of the multiple first moment from business data characteristics vector
Business data characteristics, and corresponding user's Figure Characteristics of the multiple first moment are obtained from user's Figure Characteristics vector;
Target signature determining module, for corresponding according to the initially hiding feature, the multiple first moment
Business datum feature and corresponding user's Figure Characteristics of the multiple first moment, determine the target network model
The target in duration, which is updated, in the feature hides feature;The target hides feature and pulls request with described first for characterizing
User's portrait information of corresponding target user.
Wherein, it is to pull the corresponding first request time stamp of request by described first to draw with second that the feature, which updates duration,
It takes determined by the corresponding second request time stamp of request;Described second pulls the upper drawing that request pulls request for described first
Request is taken, described second pulls request for obtaining the multiple history recommending data information for having incidence relation with target user;
Described device further include:
Relationship determination module is triggered, for updating in duration in the feature, by the multiple history recommending data information
In history recommending data information between the target user there are data-triggered relationship be determined as target histories recommending data
Information;
Traffic vector constructs module, for being based on each target histories recommending data information and the history service data
Information constructs the corresponding business datum feature vector of the target network model;It include institute in the business datum feature vector
State corresponding business datum feature of multiple first moment;The corresponding target histories recommending data of one business datum feature
Information;
User vector constructs module, for passing through the corresponding target user's category of each target histories recommending data information
Property information and the corresponding historical user's attribute information of the history service data information, construct the target network model pair
The user's Figure Characteristics vector answered;It include corresponding user of the multiple first moment in user's Figure Characteristics vector
Figure Characteristics.
Wherein, the traffic vector building module includes:
Data adding unit, for each target histories recommending data information to be added to the first business trigger list, and
The first business trigger list of each target histories recommending data information will be carried as the second business trigger list;It is described
Comprising there are history triggerings between the target user in first business trigger list and the second business trigger list
The history service data information of relationship;
Target service determination unit, in the second business trigger list, the feature to be updated in duration
Each first moment, corresponding target histories recommending data information was determined as target service data information;
Feature extraction unit is extracted for the semantic feature to the target service data information, and will be extracted
Semantic feature as the corresponding target service data characteristics of the target service data information;
Traffic vector determination unit, for obtaining the corresponding historical data feature vector of the first business trigger list,
And the target service data characteristics and the historical data feature vector are spliced, and spliced historical data is special
Sign vector is determined as the corresponding business datum feature vector of the target network model.
Wherein, the user vector building module includes:
Objective attribute target attribute determination unit is used in the second business trigger list, by the mesh at each first moment
Customer attribute information corresponding to mark history recommending data information is determined as target user's attribute information;
Feature assembled unit, for by target user's attribute information user property feature and the target network
Attention model feature in model is combined, and the user property feature after combination is determined as target user's attribute
The corresponding target user's Figure Characteristics of information;
User vector determination unit, for obtain associated with the first business trigger list historical user's feature to
Amount, and target user's Figure Characteristics and historical user's feature vector are spliced, and spliced history is used
Family feature vector is determined as the corresponding user's Figure Characteristics vector of the target network model.
Wherein, the multiple first moment includes historical juncture and object time, and the multiple first moment respectively corresponds
Business datum feature in include the historical juncture corresponding second business datum feature, the object time corresponding first
Business datum feature;Include the historical juncture corresponding the in the multiple first moment corresponding user's Figure Characteristics
The corresponding first user Figure Characteristics of two user's Figure Characteristics, the object time;The historical juncture is the object time
Last moment;
The target signature determining module includes:
History feature determination unit, for by the target network model by second user Figure Characteristics, described
Second business datum feature and the initially hiding feature are merged, and using fused initially hiding feature as described in
History of the target network model in the historical juncture hides feature;
Target signature determination unit, for by the target network model by the first user Figure Characteristics, it is described
First business datum feature and the history are hidden feature and are merged, and fused history is hidden feature as described in
Target network model hides feature in the target of the object time.
Wherein, the target network model includes the first resetting door, the second resetting door and updates door;
The target signature determination unit includes:
Weight Acquisition subelement, for it is corresponding to obtain the first business datum feature in the target network model
The first weight parameter, corresponding second weight parameter of the first user data feature, the history hide the corresponding third of feature
The straggling parameter of weight parameter and the target network model;
Parameter obtains subelement, for using the first business datum feature, first weight parameter, described first
Family Figure Characteristics, second weight parameter, the history hide feature, the third weight parameter and the straggling parameter
As the corresponding input parameter of activation primitive in the target network model, export the corresponding gating parameter of the update door,
The corresponding gating parameter of first resetting door and the corresponding gating parameter of the second resetting door;
First determines subelement, for hiding feature based on the corresponding gating parameter of the first resetting door, the history
And the first business datum feature, the corresponding time of the first business datum feature is determined in the target network model
It selects and hides feature;
Second determines subelement, for hiding feature based on the corresponding gating parameter of the second resetting door, the history
And the first user Figure Characteristics, the corresponding time of the first user Figure Characteristics is determined in the target network model
It selects and hides feature;
Third determines subelement, for according to the corresponding gating parameter of the update door, the first business datum feature
Corresponding candidate hiding feature, the corresponding candidate hiding feature of the first user Figure Characteristics and the history hide spy
Sign determines that the target network model hides feature in the target of the object time.
Wherein, described device further include:
Data acquisition module draws a portrait information from business number for hiding the user that is characterized of feature by the target
According to obtaining multiple first candidate data information in library;
First screening module, for obtaining the target user corresponding preliminary screening rule, and from the multiple first
The first candidate data information that candidate data information sifting is consistent with the preliminary screening rule is as the second candidate data information;
Second screening module, for the selection from all second candidate data information filtered out and the hiding spy of the target
Levy the user's Figure Characteristics vector characterized, the second candidate data information conduct that the business datum feature vector is consistent
Data information to be recommended;
Data distributing module, for being handed down to institute for selected data information to be recommended as target recommending data information
State target user.
Wherein, the data distributing module includes:
Probability prediction unit, for predicting selected multiple data informations to be recommended by the first clicking rate prediction model
Each of the corresponding click probability of data information to be recommended;
Sequencing unit, for being ranked up to the corresponding click probability of each data information to be recommended, and from sequence
Target recommending data information is obtained in data information to be recommended afterwards, and gives the target recommending data information recommendation to the mesh
Mark user.
Wherein, the sequencing unit is specifically used for corresponding to each data information to be recommended based on drop ordering rule
Click probability be ranked up, and screening is clicked probability and is greater than and clicks probability threshold value from the data information to be recommended after sequence
The target recommending data information filtered out is pushed to target and used by data information to be recommended as target recommending data information
Family.
Wherein, the target recommending data information includes multiple sub- recommending data information;
The data distributing module further include:
Predicted vector generation unit, for according to every sub- recommending data information in the multiple sub- recommending data information
Click probability generate the corresponding predicted vector of the first clicking rate prediction model;The predicted vector include target prediction to
Amount;The target prediction vector pulls the subsequent time of request to the multiple described first for describing the target user
The prediction probability that the first sub- recommending data information in sub- recommending data information is clicked;
Vector generation unit is clicked, the target user is recorded in the subsequent time for the multiple son for working as
When the click relationship of the second sub- recommending data information in recommending data information, generate at described first point based on the click relationship
Hit the corresponding click vector of rate prediction model;The click vector includes the target point amount of hitting to;The target point amount of hitting to is used for
The target user is described in the subsequent time to the second sub- recommending data information in the multiple sub- recommending data information
The actual probabilities clicked;
Optimize subelement, determines that described first clicks for the target point amount of hitting to according to the target prediction vector sum
The loss function of rate prediction model, and the loss function being optimized by gradient decreasing function, and by the damage after optimization
It loses function and is determined as objective function, and the corresponding first clicking rate prediction model of the objective function is determined as the second clicking rate
Prediction model.
On the one hand the embodiment of the present invention provides a kind of user's representation data processing unit, described device takes applied to business
Business device, described device includes: processor, memory and network interface;
The processor is connected with memory, network interface, wherein network interface is for providing data communication function, institute
Memory is stated for storing program code, the processor is for calling said program code, to execute such as the embodiment of the present invention
Method in one side.
On the one hand the embodiment of the present invention provides a kind of computer storage medium, the computer storage medium is stored with meter
Calculation machine program, the computer program include program instruction, execute such as this hair when the processor executes described program instruction
Method in bright embodiment one side.
In embodiments of the present invention, response first pulls request for target application, obtains this and first pulls and request pair
The feature answered updates duration, this feature when updating a length of this first pull request and upper one and pull duration between request, the spy
Sign, which updates, may include multiple first moment in duration, the first moment of each of the multiple first moment is corresponding corresponding
Business datum feature and user's Figure Characteristics;So can be by target network model obtained by a upper feature updates in duration
Hiding feature be determined as initially hiding feature, so as to will initially hide feature, described more using the target network model
A first moment corresponding first user Figure Characteristics and corresponding first business number of the multiple first moment
It is merged according to feature, is exported so that the target network model can update in duration in this feature for describing target user's
The target of user's portrait information hides feature.Wherein, it can establish in continuous time period and (i.e. should by the target network model
Feature update duration and a upper feature update duration) user's Figure Characteristics and business datum feature between incidence relation, i.e.,
The target of target network model output, which hides feature, can combine the user under the different moments in continuous time period
Figure Characteristics and business datum feature are more bonded target use so as to accurately depict in business datum recommender system
User's portrait information of family hobby, and then can effectively ensure that the accuracy and completeness of user's representation data.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is a kind of structural schematic diagram of network architecture provided in an embodiment of the present invention;
Fig. 2 a and Fig. 2 b are a kind of schematic diagram of a scenario of data interaction provided in an embodiment of the present invention;
Fig. 3 is a kind of flow diagram of user's representation data processing method provided in an embodiment of the present invention;
Fig. 4 is the schematic diagram of a scenario of time shaft corresponding to a kind of target duration provided in an embodiment of the present invention;
Fig. 5 is a kind of schematic diagram for recording data-triggered relationship provided in an embodiment of the present invention;
Fig. 6 is the structural schematic diagram of a gating cycle neural network provided in an embodiment of the present invention;
Fig. 7 is a kind of schematic diagram of a scenario for obtaining target and hiding feature provided in an embodiment of the present invention;
Fig. 8 is the flow diagram of another user's representation data processing method provided in an embodiment of the present invention;
Fig. 9 is a kind of schematic diagram of a scenario for constructing user's Figure Characteristics vector provided in an embodiment of the present invention;
Figure 10 is a kind of schematic diagram of a scenario for obtaining target recommending data information provided in an embodiment of the present invention;
Figure 11 is a kind of schematic diagram of business datum recommender system provided in an embodiment of the present invention;
Figure 12 is a kind of structural schematic diagram of user's representation data processing unit provided in an embodiment of the present invention;
Figure 13 is the structural schematic diagram of another user's representation data processing unit provided in an embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
It referring to Figure 1, is a kind of structural schematic diagram of network architecture provided in an embodiment of the present invention.As shown in Figure 1, described
The network architecture may include service server 2000 and user terminal cluster, and the user terminal cluster may include multiple users
Terminal, as shown in Figure 1, can specifically include user terminal 3000a, user terminal 3000b, user terminal 3000c ..., user
Terminal 3000n;As shown in Figure 1, user terminal 3000a, user terminal 3000b, user terminal 3000c ..., user terminal
3000n can be connected to the network with the service server 2000 respectively, in order to which each user terminal can pass through the net
Data interaction is carried out between network connection and service server 2000.
As shown in Figure 1, each user terminal in the user terminal cluster can integrate and be equipped with target application, when this
When target application is run in each user terminal, it can be counted between above-mentioned service server 2000 shown in FIG. 1 respectively
According to interaction.Wherein, which may include voice applications, Video Applications, social application, entertainment applications etc.;In consideration of it, this
The internet informations such as song, film, the article with the target application with incidence relation can be referred to as industry by inventive embodiments
Business data information.
For ease of understanding, the embodiment of the present invention is integrated with the society so that the target application is social application as an example to illustrate
Hand over the user terminal of application by realizing between the corresponding business datum platform of the target application and the service server 2000
The detailed process of user data interaction.Further, Fig. 2 a and Fig. 2 b is referred to, is a kind of data provided in an embodiment of the present invention
Interactive schematic diagram of a scenario.Target terminal user as shown in Figure 2 a can be in above-mentioned user terminal cluster shown in FIG. 1
User terminal 3000a.In addition, target user as shown in Figure 2 a can be understood as registering the social activity by target transitional information and answer
User A.As shown in Figure 2 a, user A can be at the M1 moment, by the corresponding business datum platform of the social application to figure
Service server shown in 2a (i.e. above-mentioned service server 2000 shown in FIG. 1) requests at least one service data information (institute
Stating at least one service data information may include the internet informations such as article, song, video), and can be at current time
When the M2 moment by target terminal user request described at least one service data information be referred to as target recommending data letter
Breath.It should be appreciated that the M1 moment (initiating the first drawing to go at the time of request) in the embodiment of the present invention (receives with the M2 moment
At the time of target recommending data information) between can also be spaced one or more moment.Between the M1 moment and M2 moment
Duration is the response time that above-mentioned service server 2000 pulls request for above-mentioned first, the business within the response time
Server 2000 can hide feature based on the obtained target of target network model and filter out from service database and the mesh
The target recommending data information that user's portrait information of mark user is consistent carries out data recommendation.
It can be seen that when user A (i.e. target user) there are clickstream data behaviour between above-mentioned M1 moment and target application
When making (for example, page furbishing shown in Fig. 2 a operates), background server (i.e. industry described in Fig. 2 a corresponding to the target application
Business server) it can receive the target terminal user transmission first based on the operation of aforementioned page furbishing for being integrated with the target application
It pulls request, and target account information entrained in request first can be pulled (for example, logging in target application when institute according to this
The user account information etc. used), target network corresponding with the target account information is obtained in business datum recommender system
Model.It can be used for characterizing before the M1 moment it should be appreciated that the target of target network model output hides feature
Target user user draw a portrait information, at this point, the target hide feature can be used for describing the user in continuous time period picture
The relationship changed as the time in feature and business datum feature at any time sequence.It should be appreciated that the business datum recommends system
System user's Figure Characteristics according to associated by each time in time series construct to obtain user's Figure Characteristics vector, may be used also
Business datum feature vector is obtained with the business datum feature construction according to associated by each time in the time series.Wherein,
User's Figure Characteristics vector can be used for describing each user's Figure Characteristics with the time in aforesaid time sequence drilling for changing
Change relationship.Wherein, business datum feature vector can be used for describing each business datum feature with the time in aforesaid time sequence
And the differentiation relationship changed.
For example, service server shown in Fig. 2 a, which can respond above-mentioned first, pulls request, obtains this and first pull request
The feature between request (i.e. previous page refresh operation corresponding second pulls request), which is pulled, with the last time updates duration.At this
First, which pulls the corresponding feature of request, updates in duration, and the target which is currently exported hides feature can be simultaneous
The business datum feature cared in all service data informations (for example, article information) that the target user consulted becomes at any time
The differentiation relationship of change can also be taken into account and user's Figure Characteristics in customer attribute information corresponding with each service data information
The differentiation relationship changed over time.Wherein, it should be understood that this second pulls request and can be used for asking to above-mentioned service server
The service data information for asking user's portrait information of target user when pulling request with second to be consistent.For convenient for being pulled with first
The multiple service data informations (i.e. above-mentioned target recommending data information or multiple recommending data information) requested carry out
Difference, the embodiment of the present invention, which can pull the multiple service data informations requested by second and be referred to as multiple history, to be pushed away
Recommend data information, i.e., above-mentioned service server can by the multiple history recommending data information be handed down to the target user it
Afterwards, the triggering relationship between the target user and each history recommending data information is recorded.So as to get first drawing
The corresponding feature of request is taken to update in duration, by there are data between target user in the multiple history recommending data information
The history recommending data information of triggering relationship can be referred to as target histories recommending data information, it can by the target user institute
The article information actually read is referred to as target histories recommending data information.
Wherein, which can be updated all article informations consulted in duration in feature and pressed by service server
Obtained sequence is referred to as business datum reading sequence after being arranged according to stamp of corresponding triggered time, which reads
Sequence is to pull the sequence that the article information successively consulted between request is constituted twice at this.I.e. this feature updates duration
It inside may include multiple triggered time stamps, each triggered time stamp can be understood as a moment in time series.At this
Referred to as the first moment at the time of being updated this feature in duration in inventive embodiments, and request pair can be pulled by second
The upper feature answered updates referred to as the second moment at the time of in duration.Wherein it is possible to understand, feature shown in Fig. 2 b is more
New duration is to pull the corresponding first request time stamp of request by first to pull the corresponding second request time stamp of request with second
It is identified.This feature, which updates the stamp of multiple triggered times in duration, to be the stamp of triggered time shown in Fig. 2 b 1, triggered time stamp
2, the triggered time stamp 3 ..., the triggered time stab n.As shown in Figure 2 b, each triggered time stamp with corresponding user's Figure Characteristics
There are corresponding relationships between business datum feature.For example, triggered time stamp 1 can correspond on the time shaft t shown in Fig. 2 b
The portrait of user shown in Fig. 2 b 1 and business datum 1;Triggered time stamp 2 can be with the portrait of user shown in corresponding diagram 2b 2 and business number
According to 2;And so on, the triggered time stabs n can be with the portrait of user shown in corresponding diagram 2b n and business datum n.In the embodiment of the present invention
In can by hidden state 1, hidden state 2 shown in Fig. 2 b, hidden state 3 ... hidden state n is referred to as net shown in Fig. 2 b
Network model (i.e. above-mentioned target network model) obtained hiding feature, and user shown in Fig. 2 b can be drawn a portrait 1, Yong Huhua
As 2, user's portrait 3 ..., the user n that draws a portrait be referred to as above-mentioned user's Figure Characteristics, and can by business datum 1 shown in Fig. 2 b,
Business datum 2, business datum 3 ..., business datum n be referred to as above-mentioned business datum feature.
Wherein, it should be understood that this feature updates article information (each business datum i.e. shown in Fig. 2 consulted in duration
Service data information corresponding to feature) it can be above-mentioned target histories recommending data information, i.e. the target histories recommending data
Information can be for there are the service data informations of data-triggered relationship between user A, and the target histories recommending data information
A history in the multiple history recommending data information for requesting to be issued, which is pulled, for second for the service server recommends number
It is believed that breath.It can be seen that this feature updates multiple first moment in duration, by between the target user, there are data touchings
The stamp determination of triggered time corresponding to the target histories recommending data information of hair relationship.In consideration of it, business shown in Fig. 2 a takes
Business device can find corresponding industry of each first moment from the user's Figure Characteristics vector sum business datum feature vector constructed
It is engaged in data characteristics and user's Figure Characteristics, and a upper feature can be updated in duration in obtained multiple hiding features most
The latter hides feature and is referred to as initially hiding feature, so that target network model (for example, Recognition with Recurrent Neural Network model) is being schemed
Triggered time shown in 2b exports above-mentioned hidden state n when stabbing n, and hidden state n can be referred to as to target and hide feature,
User's Figure Characteristics information of target user when so as to pull request based on the target hides that feature characterized first
Data screening, and the data information to be recommended (recommendation shown in Fig. 2 that will be filtered out are carried out in the service database shown in Fig. 2 a
Data 1, recommending data 2, recommending data 3 ....Recommending data m) is carried out online pre- by clicking rate prediction model shown in Fig. 2 b
It surveys, so as to obtain the probability of m (for example, 20) data informations to be recommended, so as in m data informations to be recommended
It is middle using probability value higher k (for example, probability more forward 6) data informations to be recommended as under target recommending data information
The target user is issued, so that the corresponding target terminal user of the target user can be at the above-mentioned M2 moment to the user A
It shows received target recommending data information, so as to ensure the accuracy of data recommendation, and the social activity can be enhanced
The viscosity of the target user in.Wherein, original state shown in Fig. 2 b can be referred to as initially to hide feature, if user A
For the old user in the social application, then it can be the network model (i.e. target network model) upper one that this, which initially hides feature,
Feature updates the obtained hiding feature of duration, it should be understood that the target network model updates in duration in a upper feature to be obtained
The detailed process of the hiding feature at corresponding second moment, can be together referring to multiple shown in Fig. 2 b to obtaining in the embodiment of the present invention
The description for hiding feature, will not continue to repeat here.Optionally, if user A is the use in the new registration social application
Family, then the initial therefore feature can be hiding feature when initializing to the network model.Wherein, m, n, k can be with
For the positive integer greater than 0.It should be appreciated that target in the embodiment of the present invention hides feature and is also used as this and first pulls request
Next the initial of target network model when pulling request (i.e. third pulls request) hide feature.
Wherein, above-mentioned continuous time period can be referred to as target duration by the embodiment of the present invention, i.e., can in the target duration
To update multiple second moment in duration comprising an above-mentioned upper feature, can also include multiple in features described above update duration
First moment.Wherein, user's Figure Characteristics vector can be used for describing corresponding to the customer attribute information in the target duration
The changing features relationship that user's Figure Characteristics change over time, business datum feature vector can be used for describing in the target duration
Service data information corresponding to the changing features relationship that changes over time of business datum feature.Wherein, target duration can be with
It is registered for target user and first pulls duration corresponding to this time interval of request to initiating this after the target application, change speech
It, may include above-mentioned first and pull the corresponding feature update duration of request, can also pull comprising second in the target duration
A corresponding upper feature is requested to update duration.
For ease of understanding, the embodiment of the present invention updates two moment in duration (when i.e. above-mentioned Fig. 2 b institute n is 2 with feature
Triggered time stamp 1 and the triggered time stamp 2) for, be illustrated with user's Figure Characteristics to the two different moments.Its
In, can will stab 1 in the triggered time corresponding at the time of referred to as k1 moment, and corresponding to stabbing 2 in triggered time at the time of, is referred to as
For the k2 moment.Customer attribute information of the user A (i.e. target user) at the k1 moment may include multiple sub- attribute informations, described more
A sub- attribute information can be with are as follows: gender: male, the age 50, region: Jiangsu, user tag: classification: sketch, tomato, anti-cancer are supported
Raw health, cuisines;After the customer attribute information at the k1 moment is input to vector embeding layer, every height category can be extracted
The property corresponding user property feature of information, and then these user property features are given to from attention layer, so as to be somebody's turn to do
User Figure Characteristics (i.e. above-mentioned user portrait 1) of the user A at the k1 moment.Further, when the target user rides on a train the Congjiang
Soviet Union has gone to Shanghai, then can describe the customer attribute information at the k2 moment behind arrival Shanghai are as follows: gender: male, age 50, ground
Domain: Shanghai, user tag: sketch, tomato, anti-cancer, classification: health-preserving, cuisines;Similarly, which can incite somebody to action
The customer attribute information at the k2 moment be given to respectively vector embeding layer and from after attention layer also the available user A in k2
User's Figure Characteristics (i.e. above-mentioned user's portrait 2) at moment.Since the target user is endless in the customer attribute information at k1 moment
It is entirely identical in the customer attribute information at k2 moment as the target user, in order to improve the data recommendation of business datum recommender system
Accuracy user's Figure Characteristics can model corresponding to the customer attribute information to different moments, to obtain to retouch
The corresponding user's Figure Characteristics vector of the target user is stated, and the target user can be described with user's Figure Characteristics vector
User draw a portrait evolution information.It therefore, can after having incorporated user portrait evolution information in business datum recommender system
It, can also be with the use at Shanghai to ensure that recommended data recommendation result can be related to user's Figure Characteristics at Wuhan
Family Figure Characteristics are related.
In consideration of it, being used to describe user's picture by splicing user's Figure Characteristics of different moments to construct to obtain
As user's Figure Characteristics vector of evolution information.It will be related duration can will to be updated to this feature in user's Figure Characteristics vector
Customer attribute information corresponding to each user's Figure Characteristics of connection is referred to as user's portrait evolution sequence, user portrait evolution sequence
A customer attribute information in column, which can be read with above-mentioned business datum between a service data information in sequence, to be existed
Corresponding relationship.In other words, multiple and different moment be may include in user's Figure Characteristics vector (for example, for example, above-mentioned Fig. 2 b
Shown in the stamp of triggered time corresponding to multiple business datum features) corresponding user's Figure Characteristics.
Wherein, it should be understood that in embodiments of the present invention, which can be obtained hiding by the k1 moment
Feature is by the input feature vector as the above-mentioned k2 moment, at the same time, in embodiments of the present invention can also be by the user at k2 moment
The business datum feature at Figure Characteristics and k2 moment is used as the input feature vector of the Recognition with Recurrent Neural Network together, i.e., at this point, the circulation
Neural network can further pass through gating cycle unit (for example, XGRU gating cycle neural network) for the input in multiple dimensions
Feature is merged, and hides feature to obtain target corresponding to the k2 moment.Wherein, the embodiment of the present invention can be by this feature more
There is the k2 moment of maximum triggered time stamp to be referred to as object time, and can be by the k1 moment before the k2 moment in new duration
The referred to as historical juncture hides so as to which the Recognition with Recurrent Neural Network is referred to as history in k1 moment obtained hiding feature
Feature.Hiding feature due to the target of Recognition with Recurrent Neural Network model output can be used for characterizing the target user when different
Therefore user's Figure Characteristics and business datum feature at quarter hide spy by the target that the target network model is currently exported
Sign can effectively ensure that the accuracy and completeness of portrayed user's portrait information.
Wherein it is possible to understand, the Recognition with Recurrent Neural Network corresponding to the present invention in embodiment is mainly used for from upper one
It is secondary when recalling finish time (i.e. second pulls request corresponding second request time stamp) to this recall request of initiation at the time of
The business of all target histories recommending data information during (i.e. above-mentioned first pulls the corresponding first request time stamp of request)
Data characteristics extracts, and the time stabbed so as to these business datum features for extracting these according to the triggered time is suitable
After sequence is successively spliced to historical data feature vector, with building obtain the corresponding business datum feature of the Recognition with Recurrent Neural Network to
Amount.In other words, the business datum feature vector may include multiple and different moment (i.e. multiple and different click timestamps) respectively
Corresponding business datum feature.Wherein, the historical data feature vector is by there are historical datas between the target user
Determined by the business datum feature of the history service data information of triggering relationship.
Wherein, service server obtains feature and updates duration, initial hiding feature and determine that target hides the tool of feature
Body process may refer to embodiment corresponding to following Fig. 3 to Figure 11.
Further, Fig. 3 is referred to, is a kind of process of user's representation data processing method provided in an embodiment of the present invention
Schematic diagram.As shown in figure 3, the method includes at least:
Step S101, response pull request for the first of target application, obtain described first and pull the corresponding spy of request
Sign updates duration;
Specifically, user's representation data processing unit, which can be responded, pulls request for the first of target application, will be described
First, which pulls request corresponding upper one, pulls request and is determined as second and pull request;Further, it pulls and asks according to described second
It asks corresponding second request time stamp and first to pull the corresponding first request time stamp of request, can determine that described first pulls
Corresponding feature is requested to update duration;
Wherein, described second multiple history recommendation that request can be used for acquisition and target user has relationship relationship is pulled
Data information, so, user's representation data processing unit can pull the corresponding feature update duration of request getting first
When, by the history recommending data information between target user there are data-triggered relationship from multiple history recommending data information
When being determined as target histories recommending data information, and then each target histories can be recommended to the corresponding triggering of attribute information
Between stamp be referred to as a moment.In other words, this feature, which updates, may include multiple moment in duration, and this feature updates in duration
Each moment may be collectively referred to as the first moment.
Wherein, user's representation data processing unit is interpreted as a kind of equipment with data screening function, the user
Representation data processing unit can integrate in destination server, which can be embodiment corresponding to above-mentioned Fig. 1
In service server 2000.Wherein, destination server can be applied in business datum recommender system, for receiving mesh
When pulling request transmitted by mark user terminal, it can be taken according to the hiding feature that target network model exports to the target
There is business device the target terminal user of network connection relation to push filtered out service data information.
Wherein it is possible to understand, the request that pulls in embodiments of the present invention may include first and pull request, second
It pulls request and third pulls request, the 4th pulls request.Wherein, it the described 4th pulls request and pulls request for described second
Upper one pulls request, i.e., the 4th pulls the corresponding 4th request time stamp of request will pull request corresponding second earlier than second
Request time stamp;Similarly, described second pull request and pull request for described first and corresponding upper one pull request, i.e., this second
The corresponding first request time stamp of request will be pulled earlier than first by pulling the corresponding second request time stamp of request.Wherein, described
Third pull request for first pull request it is corresponding it is next pull request, i.e., the third pulls the corresponding third of request and requests
Timestamp will be later than first and pull the corresponding first request time stamp of request.In embodiments of the present invention, destination server can be with
The period that two of arbitrary neighborhood pull between request is referred to as target signature and updates duration.For example, implementing in the present invention
The target signature update duration between the second request time stamp and the first request time stamp can be referred to as first in example to pull
Corresponding feature is requested to update duration.
It should be appreciated that the 4th request time can also be stabbed between second request time stamp in embodiments of the present invention
Target signature update duration and be referred to as second and pull the corresponding upper feature update duration of request;It similarly, can will be by first
Target signature between request time stamp and third request time stamp updates duration and is referred to as third to pull request corresponding next
Feature updates duration.
In consideration of it, when destination server gets and pulls request for the first of the target application, it is available this
One pulls the corresponding first request time stamp of request, and available second pulls the corresponding second request time stamp of request, from
And request can be pulled so that this based on the first request time stamp and the second request time stamp, is calculated (i.e. first pulls and ask
Ask) corresponding feature updates duration.
In addition, in embodiments of the present invention, when which can also be registered the target application by destination server
Time stab is registion time stamp, and can will be stabbed to this from the registion time and pull the corresponding first request time stamp of request
Between period be referred to as target duration, i.e., the target duration can be understood as registering from the target user target application to
It initiates this and pulls duration corresponding to the time interval between requesting.Therefore, it may include in the target duration comprising multiple
Target signature updates duration, first pulls the corresponding feature of request for example, may include this and updates duration, can also include second
It pulls the corresponding upper feature of request and updates duration etc..In consideration of it, in embodiments of the present invention, request can be pulled by first
(i.e. this pulls request) and second pulls between request (i.e. upper one pulls request) target signature and updates duration and be referred to as first
Feature updates duration, and fisrt feature update duration can be referred to as to first and pull the corresponding feature update duration of request;
Similarly, can by second pull request and the 4th pull request between target signature update duration be referred to as second feature more
New duration, and second feature update duration can be referred to as to second and pull the corresponding upper feature update duration of request.It changes
Yan Zhi, which updates duration and second feature update duration is two continuous durations in the target duration.It answers
Work as understanding, each target signature, which updates in duration, may include multiple and different moment, and for ease of understanding, the embodiment of the present invention can
Each moment in above-mentioned upper feature update duration was referred to as the second moment, and features described above can be updated in duration
Each moment be referred to as the first moment.Wherein, multiple first moment that this feature updates in duration may include the historical juncture
And object time.Wherein, at the time of the object time can update corresponding with stamp of maximum triggered time in duration for this feature.
It should be appreciated that for convenient for updated with this feature in the embodiment of the present invention have in duration maximum triggered time stamp when
Quarter is distinguished, and the embodiment of the present invention will can also also be referred to as at the time of having maximum triggered time stamp in a upper renewable time
It is initial hiding special so as to which target network model to be referred to as in the obtained hiding feature of the initial time for initial time
Sign.
Wherein, it should be understood that before executing step S101, destination server can respond second for target application
Request is pulled, the initial feature of hiding based on target network model is screened with the target user from service database with pass
Multiple history recommending data information of connection relationship, and the multiple history recommending data information is being pushed to the target user
Later, the data-triggered relationship between the target user and multiple history recommending data information is recorded.And then it may insure this
Destination server the feature update duration in, by the multiple history recommending data information between the target user
There are the history recommending data information of data-triggered relationship to be determined as target histories recommending data information;Further, based on every
A target histories recommending data information and history service data information construct the corresponding business datum of the target network model
Feature vector;It include corresponding business datum feature of the multiple first moment in the business datum feature vector;Into
One step, pass through each corresponding target user's attribute information of target histories recommending data information and the history industry
The corresponding historical user's attribute information of business data information, constructs the corresponding user's Figure Characteristics vector of the target network model;
It include corresponding user's Figure Characteristics of the multiple first moment in user's Figure Characteristics vector.Wherein, described to go through
History service data information can be understood as using before second pulls the corresponding second request time stamp of request with the target
There are all service data informations of history triggering relationship between family.At this point, all before the second request time stamp
Service data information is respectively positioned in the first trigger list, i.e., first trigger list can be used for storing all and target user
Between with data-triggered relationship service data information.
For ease of understanding, when the embodiment of the present invention only pulls request corresponding feature update with first in target duration
For length, with illustrate this second pull request and first pull request between incidence relation.Further, Fig. 4 is referred to, is
The schematic diagram of a scenario of time shaft corresponding to a kind of target duration provided in an embodiment of the present invention.Target duration as shown in Figure 4
It may include in corresponding time shaft and pull corresponding second request time of request with second and stab, second request time stamp can
Think the A1 moment shown in Fig. 4.In addition, time shaft as shown in Figure 4 can also be comprising pulling request corresponding first with first
Request time stamp, first request time stamp can be the A5 moment shown in Fig. 4.It is understood that recommending in business datum
In system, this can be pulled to the time interval (time between A1 moment and A5 moment i.e. shown in Fig. 4 between request twice
Section) be referred to as to be characterized update duration, i.e., this feature, which updates duration, can be understood as described mesh in embodiments of the present invention
Mark a period in duration.May exist multiple crucial moments shown in Fig. 4 in target duration, i.e., when multiple crucial
Quarter may include A1 moment, A2 moment, A3 moment, A4 moment, A5 moment etc..On time shaft shown in Fig. 4, the A1 moment is mesh
At the time of when mark user terminal pulls request (i.e. second pulls request) to destination server initiation;The A2 moment is to be based on second
Pull request in service database carry out data screening at the time of;The A3 moment is to return to the service data information filtered out
At the time of target terminal user, the A4 moment be target user to being triggered in received service data information at the time of,
For example, the A4 moment can for target user in target application to received service data information (for example, song, text
The internet informations such as chapter, film) it carries out click play or clicks the corresponding triggered time stamp of the trigger actions such as access;The A5 moment
At the time of when pulling request to destination server initiation again for target terminal user, at this point, this initiate again pull request
It can be referred to as first and pull request.In other words.In embodiments of the present invention, destination server can according to Fig.4,
Two, which pull the corresponding request time stamp (i.e. the second request time stamp) of request and first, pulls the corresponding request time stamp of request (i.e.
First request time stamp), determine that first pulls the corresponding feature update duration of request.This feature updates the A4 moment in duration
It can be the corresponding triggered time stamp of a target histories recommending data information.The target histories recommending data information can be figure
There is in multiple history recommending data information shown in 4 with target user the service data information of data-triggered relationship.
For example, with the 4 service data informations (business i.e. shown in Fig. 4 shown in target terminal user shown in Fig. 4
Data information 1, service data information 2, service data information 3 and service data information 4) for, when target user is to this 4 industry
When a service data information (for example, service data information 1) being engaged in data information executes trigger action 1, the target user
Terminal can respond the corresponding trigger action 1 of service data information 1, which is stabbed and is added
It is sent to destination server into clickstream data instruction, and by clickstream data instruction, so that the destination server is based on the number
The data-triggered relationship between the target user and the service data information 1 is recorded according to click commands.Similarly, as target user couple
When service data information 2 in this 4 service data informations executes trigger action 2, which can respond the industry
Stamp of the corresponding triggered time of trigger action 2 is added in clickstream data instruction by the business corresponding trigger action 2 of data information 2,
And clickstream data instruction is sent to destination server, so that the destination server is based on clickstream data instruction and records the mesh
Mark the data-triggered relationship between user and the service data information 1.And so on, which can also get
When the corresponding clickstream data of service data information 3 instructs, the data touching between the target user and the service data information 3 is recorded
Hair relationship;Similarly, which can also be when getting the instruction of the corresponding clickstream data of service data information 4, record
Data-triggered relationship between the target user and the service data information 4.Further, which can be based on note
The data-triggered relationship recorded, goes through target is referred to as there are the service data information of data-triggered relationship between target user
History recommending data information, and each target histories recommending data information can be added to the first business trigger list, so as to
The first business trigger list of each target histories recommending data information will be carried as the second business trigger list, after
And can by target histories recommending data information in the second business trigger list and the history service data information,
Construct the corresponding business datum feature vector of the target network model.It should be appreciated that at this point, in the business datum feature vector
It may include corresponding business datum feature of multiple and different moment.It should be appreciated that can be in the business datum feature vector
It include that there are the business datums of all service data informations of data-triggered relationship between the target user in target duration
Feature can describe business datum evolution information of the target user in target duration by these business datum features.This
When, the corresponding triggered time stamp of each target histories recommending data information may include the corresponding touching of service data information 1
Timestamp is sent out (for example, B41Moment), the corresponding triggered time of service data information 2 stamp (for example, B42Moment), business datum letter
3 corresponding triggered time stamps are ceased (for example, B43Moment), the corresponding triggered time of service data information 4 stamp (for example, B44Moment).
Wherein, the corresponding triggered time sequence of this four triggered time stamps can be B41Moment → B42Moment → B43When
Quarter → B44Moment.So the service data information 4 is right in the corresponding triggered time stamp of this four service data informations
The triggered time stamp answered can be maximum triggered time stamp, so can by the service data information 4 it is corresponding it is maximum trigger when
Between stamp be referred to as object time, and can will be corresponding to service data information 3, service data information 2, service data information 1
Triggered time stamp is referred to as the historical juncture.
For another example, still with service data information 1 shown in Fig. 4, service data information 2, service data information 3 and business number
It is believed that for breath 4, if target user, which only updates in duration in features described above, has accessed service data information 1, the destination service
Device can update the data-triggered relationship recorded between the target user and the service data information 1 in duration in this feature, and
The service data information 1 can be added in the first business trigger list according to the data-triggered relationship.In other words, twice
It pulls between request, is only added to service data information 1 in the first business trigger list, so as to by the second business trigger column
Service data information 1 in table is referred to as target histories recommending data information, and can be by the corresponding touching of service data information 1
Hair timestamp is determined as maximum triggered time stamp, at this point it is possible to referred to as target at the time of the maximum triggered time is stabbed corresponding
At the moment, at this point, this feature updates in duration, there will be no the values of historical juncture or the historical juncture can be considered as 0.
For another example, still with service data information 1 shown in Fig. 4, service data information 2, service data information 3 and business number
It is believed that for breath 4, if the destination server is also pressed after service data information 1 is added in the first business trigger list
It service data information 3 and business is successively familiar with information 4 is added in the first business trigger list according to triggered time stamp, then it can be with
Further by this, to be added to service data information 1, service data information 3, the first business trigger list of service data information 4 true
It is set to the second business trigger list.As shown in figure 4, may be embodied in above-mentioned A1 shown in Fig. 4 in the first business trigger list
Other history triggered times before moment stab associated service data information, and can will be in the first business trigger list
Service data information associated with other historical times stamp be referred to as history service data information.In other words, in the second industry
Each service data information in business trigger list can correspond to triggered time stamp, it is contemplated that target user every time can be with
A service data information is consulted, therefore, triggering corresponding to each service data information in the second business trigger list
Timestamp is different.In order to be distinguished with aforementioned history service data information, the embodiment of the present invention can be by new addition
Target histories are referred to as to service data information 1, service data information 3, the service data information 4 in the first business datum list
Recommending data information, and the corresponding triggered time stamp of the target histories recommending data information is determined as most newly added business and is gone out
Has the corresponding maximum triggered time stamp of information 4.
It can be seen that destination server as shown in Figure 4 is getting above-mentioned target terminal user shown in Fig. 4 in A1
When carving transmitted second and pulling request, target network model can be updated into a duration (i.e. above-mentioned upper feature in second feature
Update duration) the hiding feature of interior institute's final output is referred to as the first hiding feature, and then can at the A2 moment based on this
User that one hiding feature is characterized draw a portrait information (an i.e. upper feature update user's Figure Characteristics of the different moments in duration with
Business datum feature), data screening is carried out, from service database shown in Fig. 4 so as to filter out and the target user
Multiple history recommending data information with incidence relation, and can be filtered out these with the A3 moment shown in Fig. 4 multiple
History recommending data information is referred to as the service data information of screening, and these service data informations screened are returned to target
That is, user terminal should so that target terminal user can export the service data information of these screenings of display on a terminal screen
Target terminal user can show service data information 1 shown in Fig. 4, service data information 2, service data information 3, business number
It is believed that breath 4.
Wherein, the destination server by this four service data informations filtered out be handed down to the target terminal user into
When row is shown, it can synchronize and be pushed away in the data recorded in data-pushing list between the target user and this four service data informations
Send relationship.For example, the data-pushing relationship between target user and service data information 1 can be expressed as E1, target user with
Data-pushing relationship between service data information 2 can be expressed as E2, the data between target user and service data information 3
Push relationship can be expressed as E3, and the data-pushing relationship between target user and service data information 4 can be expressed as E4.It can
With understanding, at this point, on service data information 1, service data information 2, service data information 3, service data information 4 be
State multiple history recommending data information that there is relationship relationship with target user.When target user believes multiple history recommending datas
When some history recommending data information in breath carries out clicking operation (i.e. above-mentioned trigger action), which can be into
Data-triggered relationship between history recommending data information corresponding to the one-step recording target user and the clicking operation, and can
The corresponding history recommending data information of the clicking operation is referred to as target histories recommending data information.
It can be seen that target can be based on when destination server gets and pulls request for the second of target application
The initial hiding feature of network model screens the multiple history for having incidence relation with the target user from service database
Recommending data information, and after the multiple history recommending data information is pushed to target user, it records the target and uses
The data-triggered relationship between target histories recommending data information in family and multiple history recommending data information.
For ease of understanding, further, Fig. 5 is referred to, is that a kind of record data-triggered provided in an embodiment of the present invention is closed
The schematic diagram of system.In embodiments of the present invention, what which screened from service database has with target user
Related multiple history recommending data information, can for above-mentioned service data information 1 shown in Fig. 4, service data information 2,
Service data information 3, service data information 4.Therefore, when this 4 service data informations are handed down to shown in Fig. 5 by destination server
Target terminal user while be shown, can with recorded in data-pushing list shown in Fig. 5 the target user with
Data-pushing relationship between this four service data informations.Therefore, when in target user display interface 100a shown in Fig. 5
When executing clicking operation 1 to the service data information 1 in shown four service data informations, the clicking operation pair can be generated
The clickstream data instruction answered, and the clickstream data can be instructed and be sent to destination server shown in fig. 5, so that the target takes
Business device can be instructed according to the corresponding clickstream data of clicking operation 1, and the corresponding data content of service data information 1 is returned to
The target terminal user, so that the target terminal user can show received data content in the service data information
1 corresponding content presentation interface, in order to which can to consult the service data information 1 on the content presentation interface right by target user
The data content answered to ensure that the result of data recommendation can be more accurate, and then can increase the target user in the target
Duration is enlivened in.In other words, target terminal user, can when receiving 1 corresponding data content of service data information
The display interface 100a of the display interface of terminal as shown in Figure 5 is switched to foregoing teachings display interface.It is understood that
The destination server can also be synchronized when the corresponding data content of service data information 1 is returned to the target terminal user
Data-triggered relationship between the target user and the service data information 1 is added to data-pushing list shown in fig. 5.It changes
Yan Zhi, the destination server can be set in data-pushing list for the state tag in 1 place list column of service data information
It is set to click label, and the corresponding triggering of service data information 1 can be recorded in the list column where service data information 1
Timestamp (for example, A4 moment in embodiment corresponding to above-mentioned Fig. 4).
It should be appreciated that the data-pushing list can be used for recording the target user and each history recommending data of push is believed
Data-pushing relationship between breath, which can be also used for recording each history recommending data information and target is used
Data-triggered relationship between family, it can be corresponding in data-pushing list according to the data-triggered relationship being recorded
Label has been clicked in the setting of history recommending data information, so as to which going through for label has been clicked in carrying in the data-pushing list
History recommending data information is referred to as target histories recommending data information.Further, destination server can be in the data-pushing
The history recommending data information that label has been clicked in above-mentioned carrying is added in the first business trigger list in list, first industry
Business trigger list can be above-mentioned first business trigger list shown in Fig. 4.Therefore, when destination server is by mesh shown in fig. 5
Mark history recommending data information is added to after the first business trigger list, this can be added to target histories recommending data letter
First business trigger list of breath is referred to as the second business trigger list, it can obtains the second business trigger column shown in fig. 5
Table.Wherein, it may include in the second business trigger list shown in fig. 5 in above-mentioned first business trigger list shown in Fig. 4
History service data information, can also further include the service data information 1 being newly added in the first business trigger list,
Service data information 3 and service data information 4.In embodiments of the present invention, if the corresponding triggered time of service data information 1 stabs
For A41Moment, stamp of the corresponding triggered time of service data information 2 are A42Moment, stamp of the corresponding triggered time of service data information 3
For A43Moment, and the corresponding triggered time sequence of these three service data informations is A41Moment → A42Moment → A42Moment then should
Destination server, which can stab the corresponding triggered time of service data information 4, be determined as maximum triggered time and stab, and by business number
It is believed that the corresponding maximum triggered time stamp of breath 4 is referred to as object time, and can be by the corresponding triggered time of service data information 3
Stamp and stamp of the corresponding triggered time of service data information 1 are referred to as the historical juncture.
Wherein, for ease of understanding, further, table 1 is referred to, is a kind of data-pushing column provided in an embodiment of the present invention
The corresponding distribution situation table of partial service data information in table.As shown in table 1, target clothes are record in the data-pushing list
Business device is based on second and pulls be handed down to (3000a pairs of user terminal in embodiment corresponding to i.e. above-mentioned Fig. 1 of target user of request
The user A answered) multiple history recommending data information, multiple multiple history recommending data information can be right for above-mentioned Fig. 4
Answer service data information 1, service data information 2, service data information 3 and the service data information 4 in embodiment.
Table 1
As shown in Table 1 above, user A can be A in current time41When the moment, to above-mentioned display interface shown in fig. 5
Service data information 1 in 100a executes clicking operation 1, so as to obtain the corresponding clickstream data of clicking operation 1 instruction
1, so as to which the clickstream data 1 instruction is sent to destination server, so that the destination server can count shown in table 1
According to the data-triggered relationship 1 recorded between user A and the service data information 1 in push list, in other words, destination server can
With the data-triggered relationship 1 that is recorded based on this by 1 place list of service data information in the data-pushing list shown in table 1
The state tag on column is set as having clicked label, and can indicate to have clicked label with numerical value 1 in the data-pushing list.With
This simultaneously, which can also stab the number shown in table 1 according to the triggered time entrained in the data-triggered relationship 1
According to push list in by the corresponding triggered time stamp of service data information 1 be table 1 shown in A41Moment.
Similarly, user A can also be A in current time42When the moment, in above-mentioned display interface 100a shown in fig. 5
Service data information 3 executes clicking operation 2, so as to obtain the corresponding clickstream data of clicking operation 2 instruction 2, so as to
The clickstream data 2 instruction is sent to destination server, so that the destination server can the column of the data-pushing shown in table 1
The data-triggered relationship 2 between user A and the service data information 3 is recorded in table, in other words, destination server can be based on should
The data-triggered relationship 2 being recorded is in the data-pushing list shown in table 1 by the state on 3 place list column of service data information
Label is set as having clicked label, and can indicate to have clicked label with numerical value 1 in the data-pushing list.At the same time,
The destination server can also be stabbed according to the triggered time entrained in the data-triggered relationship 2, the data-pushing shown in table 1
In list by the corresponding triggered time stamp of service data information 3 be table 1 shown in A42Moment.And so on, the destination service
The corresponding triggered time stamp of service data information 4 can also be shown in table 1 in the data-pushing list shown in table 1 by device
A43Moment.So in embodiments of the present invention, it can be according to the triggering of the corresponding triggered time stamp of each service data information
Stamp of the corresponding triggered time of service data information 4 is determined as maximum triggered time and stabbed by time sequencing, so as to by business number
It is believed that 4 corresponding A of breath43Moment can be stamp of maximum triggered time corresponding to target histories recommending data information.In other words, mesh
Mark server can by state-of-the-art record to triggered time stamp be determined as target histories recommending data information corresponding triggered time and stab,
Be referred to as object time so as to update the triggered time for the state-of-the-art record being arrived in duration stamp in feature, and by A42Moment and
A41Moment is referred to as the historical juncture together.
Further, which can will carry the business datum for having clicked label from the data-pushing list
Information 1, service data information 3 and service data information 4 are referred to as target histories recommending data information, so as to carry
Service data information 1, service data information 3 and the service data information 4 for clicking label are successively added to above-mentioned shown in Fig. 4 the
In one business trigger list, so as to which the service data information 1, service data information 3 and service data information 4 will be added to
The first business trigger list be referred to as the second business trigger list, specifically, the second business trigger column shown in Figure 5
The schematic diagram of table.
Wherein it is possible to understand, which can be according to the target histories in the second business trigger list
Recommending data information obtains the user A (i.e. target user) and reads sequence in the business datum that feature updates in duration, so as to
To read the business datum feature and history service data information pair of each service data information in sequence based on the business datum
The historical data feature vector answered constructs the user A in the industry at current time (A5 moment in embodiment corresponding to i.e. above-mentioned Fig. 4)
Business data characteristics vector.Wherein it is possible to understand, it may include multiple connection periods in the business datum feature vector
The different moments corresponding business datum feature of interior (i.e. multiple target signatures update in duration);Similarly, the destination service
Device can pass through the corresponding target user's attribute information of target histories recommending data information and the history service data information
Corresponding historical user's attribute information constructs the corresponding user's Figure Characteristics vector of the target network model;Wherein it is possible to manage
It solves, may include the different moments difference in the multiple target signature update duration in user's Figure Characteristics vector
Corresponding user's Figure Characteristics.
Target network model is updated obtained hiding feature in duration in a upper feature and is determined as initially by step S102
Hide feature;
Wherein, the initially hiding feature can be the target network model obtained by a upper feature updates in duration
Hiding feature (i.e. above-mentioned first hiding feature);The initial time can update the maximum in duration for a upper feature
Triggered time stabs corresponding second moment.It should be appreciated that the target network model this second pull request it is corresponding it is upper one spy
Sign, which updates and hides feature obtained by the initial time in duration, to be hidden state n shown in above-mentioned Fig. 2 b.I.e. this is initial hidden
Hiding feature can be used for characterizing this and second pull the user's portrait information for request corresponding target user.In the embodiment of the present invention
In, historical user spy can be referred to as by constructed obtained second user Figure Characteristics vector in duration is updated in a upper characteristic
Levy vector or upper user Figure Characteristics vector, in order to the target network model can based on historical user's feature vector into
One step updates building in duration in this feature and obtains first user's Figure Characteristics vector, and by the first user Figure Characteristics vector
Referred to as user's Figure Characteristics vector.Similarly, constructed the second obtained business number in duration will can be updated in a upper characteristic
It is referred to as historical data feature vector or upper business datum feature vector according to feature vector, in order to which the target network model can
To be based on the historical data feature vector, building in duration is updated in feature and obtains the first business datum feature vector, and should
First business datum feature vector is referred to as business datum feature vector.
Wherein it is possible to understand, each user's Figure Characteristics in user's Figure Characteristics vector can be used for describing should
Customer attribute information of the target user corresponding to the different moments time in sequence and the relationship that changes at any time, and can will
The customer attribute information user that the time in sequence and the relationship that changes are referred to as the target user at any time draws a portrait evolution letter
It ceases, may include the user's portrait differentiation sequence for the target user that this feature updates in duration in user portrait evolution information
Column.It can be seen that same user (i.e. target user) can be drawn a portrait in the user of different moments by the target network model
Feature and business datum feature are merged, so that the target network model can be further to the user property of the target user
The evolution process of information is modeled, and can be modeled to the evolution process of the service data information of the target user, from
And data recommendation can be carried out according to hiding feature (for example, above-mentioned initially hiding feature) obtained after modeling, to ensure this
Second pulls corresponding recommendation results is requested second can to pull the corresponding upper feature of request with this and update industry in duration
Data of being engaged in read sequence correlation, can also combine the evolution history for the customer attribute information that a feature updates in duration.
It is special to obtain corresponding business datum of the multiple first moment from business data characteristics vector by step S103
Sign, and corresponding user's Figure Characteristics of the multiple first moment are obtained from user's Figure Characteristics vector;
Specifically, the corresponding business datum feature vector of the available target network model of destination server and user
Figure Characteristics vector;If the multiple first moment includes above-mentioned historical juncture and object time, can be drawn from the user
As obtaining the historical juncture corresponding user's Figure Characteristics in feature vector as second user Figure Characteristics, and from the industry
The historical juncture, which is obtained, in business data characteristics vector corresponds to business datum feature as the second business datum feature;Similarly, should
Destination server can also obtain the corresponding user's Figure Characteristics of the object time from user's Figure Characteristics vector and make
For first user's Figure Characteristics, and the object time is obtained from the business datum feature vector and corresponds to business datum feature
As the first business datum feature.
Wherein, the object time can update have maximum triggered time stamp corresponding in duration first for this feature
Moment, the historical juncture are that the stamp of remaining triggered time that this feature updates in duration in addition to the object time is corresponding
First moment.At the time of i.e. the object time and historical juncture are characterized in update duration.
Step S104, according to the initially hiding feature, corresponding business datum feature of the multiple first moment,
And the multiple first moment corresponding user's Figure Characteristics, determine that the target network model is updated in the feature
Target in duration hides feature;
Specifically, application server can be by the target network model by second user Figure Characteristics, described
Second business datum feature and the initially hiding feature are merged, and using fused initially hiding feature as described in
History of the target network model in the historical juncture hides feature;It further, will be described by the target network model
First user's Figure Characteristics, the first business datum feature and the history are hidden feature and are merged, and will be after fusion
History hide feature as the target network model the object time target hide feature.
For ease of understanding, the embodiment of the present invention is believed with the business datum in above-mentioned second business trigger list shown in fig. 5
For breath 1, service data information 3 and service data information 4, with the difference of multiple first moment got to the destination server
Corresponding user's Figure Characteristics and business datum feature are illustrated.Destination server can be from constructed business datum feature
A is obtained in vector41The business datum feature 1, A at moment42The business datum feature 2 and A at moment43The business datum feature at moment
3.At this point, the business datum feature vector can be expressed as X={ x1, x2..., xn, xn+1,xn+2,xn+3}.In business datum spy
It levies in vector, { x1, x2..., xnIt can be used to indicate that historical data feature vector;Wherein, xn+1,xn+2,xn+3It may be collectively referred to as
The target service data characteristics extracted from corresponding target histories recommending data information of each first moment, i.e. destination service
Device can splice target service data characteristics and historical data feature vector, and by spliced historical data feature to
Amount is determined as the corresponding business datum feature vector of the target network model.Wherein, xn+1It can be used to indicate that shown in fig. 5
The business datum feature 1, x of service data information 1n+2It can be used to indicate that the business datum of service data information 3 shown in fig. 5
Feature 2, xn+3It can be used to indicate that the business datum feature 3 of service data information 4 shown in fig. 5.Similarly, destination server can
With by (A when aforementioned different moments41Moment, A42Moment, A43Moment) user's Figure Characteristics and historical user's feature vector
(for example, { u1, u2..., un) spliced, and spliced historical user's feature vector is determined as the target network mould
The corresponding user's Figure Characteristics vector of type, user's Figure Characteristics vector can be expressed as U={ u1, u2..., un, un+1,un+2,
un+3}.Wherein, un+1It can be used to indicate that the corresponding user's Figure Characteristics 1, u of service data information 1 shown in above-mentioned table 1n+2It can
For indicating the corresponding user's Figure Characteristics 2, u of service data information 3 shown in above-mentioned table 1n+3It can be used to indicate that above-mentioned table
The corresponding user's Figure Characteristics 3 of service data information 4 shown in 1.
In consideration of it, updating in duration in feature, target network model can be carved (i.e. target by destination server at the beginning
Network model is at the time of an above-mentioned upper feature updates in duration) obtained hiding feature as the historical juncture (for example, above-mentioned table
A shown in 141Moment) it is initial hide feature (for example, h0), and described in obtaining from user's Figure Characteristics vector (i.e. U)
Historical juncture (i.e. A41Moment) corresponding user's Figure Characteristics are as second user Figure Characteristics (i.e. un+1), and from the business
Historical juncture (the i.e. A is obtained in data characteristics vector (i.e. X)41Moment) business datum feature is corresponded to as the second business datum
Feature (i.e. xn+1);Wherein, the initial time can update the maximum triggered time in duration for a feature upper in time series
Stab corresponding second moment;Further, which can be by the target network model by the second user
Figure Characteristics (i.e. un+1)), the second business datum feature (i.e. xn+1)) and initially hiding feature (i.e. the h0) melted
It closes, and using fused initially hiding feature as the target network model in historical juncture (the i.e. A41Moment) history
It hides feature (for example, h1), i.e., it can be the target network model in A that the history, which hides feature (h1),41Moment is exported hidden
Hide feature.
And so on, it is updated in duration in this feature, which can be further by target network model in A41
Moment obtained hiding feature is as the new historical juncture (for example, A shown in above-mentioned table 142Moment) it is new initial hiding
Feature (for example, h1), and new historical juncture (the i.e. A is obtained from user's Figure Characteristics vector (i.e. U)42Moment) it is corresponding
User's Figure Characteristics as new second user Figure Characteristics (i.e. un+2), and from the business datum feature vector (i.e. X)
Obtain new historical juncture (the i.e. A42Moment) business datum feature is corresponded to as the second new business datum feature (i.e. xn+2);
Wherein, the historical juncture and the new historical juncture are the first moment in this feature update duration;Further, should
Destination server can be by the target network model by new second user Figure Characteristics (the i.e. un+2)), it is described new
Second business datum feature (i.e. xn+2)) and new initial hide feature (i.e. the h1) merged, and will it is fused newly
It is initial hide feature as the target network model in new historical juncture (i.e. A42Moment) new history hide feature
It can be the target network model in A that (for example, h2), the i.e. new history, which hide feature (h2),42The hiding spy that moment is exported
Sign.Due to the embodiment of the present invention be will have the maximum triggered time stamp at the time of referred to as object time (for example, A43Moment), institute
Can update all first moment before object time in duration in feature (for example, A41Moment, A42Moment) it is referred to as
Historical juncture.
Wherein, target network model in embodiments of the present invention can be the gating cycle in Recognition with Recurrent Neural Network model
Unit, the gating cycle unit can be understood as a kind of gating cycle neural network, the gating cycle mind in the embodiment of the present invention
It can be used for handling the time upper sequence data that discontinuously but each other there is certain relationship through network.For example, with shown in above-mentioned table 1
Three triggered times stab (A41Moment, A42Moment .A43Moment) for, by the gating cycle neural network can to this three
User's Figure Characteristics and business datum feature under a different moments model, so as to preferably capture above-mentioned table 1
In each triggered time stamp between dependence.It wherein, can in the gating cycle neural network (i.e. target network model)
With by A shown in above-mentioned table 141Moment, A42Moment is referred to as the historical juncture, and by the target network model in A41Obtained by moment
To hiding feature (h1) and A42Moment obtained hiding feature (h2) is referred to as history and hides feature, and can be by above-mentioned table 1
Shown in A43Moment is referred to as object time, and the corresponding user's Figure Characteristics of the object time can be referred to as the first user
Figure Characteristics (i.e. un+3), and the corresponding business datum feature of the object time can be referred to as the first business datum feature (i.e.
xn+3).In consideration of it, sensu lato the historical juncture can be claimed by the gating cycle neural network (i.e. target network model)
The last moment for object time.
Wherein, which can modify in Recognition with Recurrent Neural Network and hide by introducing the concept of door
The calculation of state (hiding feature).The gating cycle neural network may include resetting door (the i.e. first resetting door and the
Two resetting doors), update door, candidate hidden state (i.e. candidate hiding feature) and hidden state.Wherein, gating cycle neural network
Resetting door (reset gate) and update door (update gate) input be current time (i.e. current time be t when
Carve) corresponding to input feature vector (i.e. the business datum feature at current time can be xtWith user's Figure Characteristics at current time
It can be ut) with last moment corresponding to hiding feature ht-1;The output of the gating cycle neural network (hides feature ht)
It can be calculated by the full articulamentum that activation primitive is sigmoid function (i.e. sigma function).Wherein, the gating cycle nerve net
Resetting door in network is mainly used in pull-in time sequence pulling the corresponding feature of request with above-mentioned first that update duration associated
Input feature vector between dependence, and update door be mainly used for it is associated with target duration all in pull-in time sequence
Dependence between input feature vector.
It should be appreciated that with the service data information 1 in above-mentioned second business trigger list shown in fig. 5, business datum letter
For breath 3, service data information 4, then the corresponding feature of the target network model, which updates in duration, may include three first
Moment, (n in embodiment corresponding to i.e. above-mentioned Fig. 2 b was 3, and the first moment corresponding to stamp of each triggered time can be above-mentioned
A in table 141Moment, A42Moment and A43Moment), so as to by these three first moment corresponding user's Figure Characteristics
The input feature vector of the target network model is referred to as with business datum feature, which can be understood as the use of target user
Family state feature.It is understood that in embodiments of the present invention, can according in the second business trigger list with this feature more
The new associated each service data information of duration, determines the first list entries of the target network model, and can should
First list entries be referred to as business datum read sequence, in order to it is subsequent can be by these business in first list entries
The corresponding business datum feature of data information is sequentially inputted to the target network model as input feature vector.Similarly, at this
In inventive embodiments, can also according to customer attribute information corresponding to aforementioned each service data information, determine the mesh
The second list entries of network model is marked, and second list entries can be referred to as to user's portrait evolution sequence, in order to
It is subsequent can be using the corresponding user's Figure Characteristics of these customer attribute informations in second list entries as another defeated
Enter feature and is sequentially inputted to the target network model.It wherein, can be above-mentioned first defeated by this for the target network model
Enter sequence and the second list entries is referred to as target list entries, so as to which each input in the target list entries is special
The corresponding triggered time stamp (for example, the triggered time corresponding to service data information stabs) of sign be referred to as first moment or
One time step, so as to obtain above three by the target network model (for example, gating cycle neural network) building
Dependence between first moment, in order to be able to current time be A43The target of the target network model is exported when the moment
Hide feature.
For ease of understanding, further, Fig. 6 is referred to, is a gating cycle nerve net provided in an embodiment of the present invention
The structural schematic diagram of network.As shown in fig. 6, may include above-mentioned first resetting door in the gating cycle neural network (i.e. shown in Fig. 6
Resetting door 1), it is above-mentioned second resetting door (resetting door 2 i.e. shown in fig. 6), above-mentioned update door, the first business datum feature correspondence
Candidate hide feature (h '), the second business datum feature is corresponding candidate hiding feature (h "), and uses for describing target
The target of user's portrait information at family hides feature (i.e. h).It can by resetting door 1 shown in fig. 6, resetting door 2 and update door
To save the data information in the target list entries being respectively input in the gating cycle neural network as much as possible, and will not be with
The variation of time and remove the data information in saved target list entries, will not be removed because of uncorrelated to prediction
Data information in the target list entries.It should be appreciated that the embodiment of the present invention can exist same user (i.e. target user)
Customer attribute information when feature updates each different moments in duration is integrated, and the corresponding use of the target user is obtained
Family portrait evolution sequence, further, it is also possible to update all business numbers for consulting the target user in duration in this feature
It is believed that breath is integrated, obtains the corresponding business datum of the target user and read sequence.
Wherein, gating cycle neural network as shown in FIG. 6 may include two resetting doors, one in the two resetting doors
The gating parameter of a resetting door (for example, resetting door 1 shown in fig. 6) can be denoted as rt, it is (i.e. shown in fig. 6 that another resets door
Resetting door 2) gating parameter can be denoted as gt.In addition, the gating parameter as shown in FIG. 6 for updating door can be denoted as zt.Such as Fig. 6
Shown, when by the gating cycle neural network can be t moment by current time two input feature vectors (input spy
Sign is the first user Figure Characteristics shown in fig. 6 and another input feature vector is the first business datum feature shown in fig. 6) and t-
The obtained hiding feature (hiding feature i.e. shown in fig. 6) at 1 moment (i.e. previous moment) is merged, when obtaining current
The output feature for t moment is carved, at this point, the hiding feature that the gating cycle neural network can be exported the t moment is referred to as
Feature is hidden for target shown in fig. 6.
Wherein, in gating cycle neural network shown in Fig. 6, target can be hidden into feature and is denoted as ht, and can incite somebody to action
History hides feature and is denoted as ht-1, and first user's Figure Characteristics of t moment can be denoted as ut, and can be by the first of t moment
Business datum feature is denoted as xt.In consideration of it, can be by output feature (the i.e. h of the gating cycle neural networkt) followed with following gate
The output calculation formula (1) of ring neural network is indicated:
ht=XGRU (xt,ut,ht-1)=(1-zt)°ht-1+zt°(h't+h”t) formula (1)
In formula (1), XGRU can be understood as a kind of operator, which can be used in multiple input feature vectors
After (the first user Figure Characteristics, the first business datum feature and history i.e. shown in fig. 6 hide feature) carries out Function Mapping
Obtain hidden state h shown in fig. 6 (hide feature), and can will be the obtained hiding feature of t moment at current time into
Row output hides feature (i.e. h to export target shown in fig. 6t).In other words, it is updated in duration in feature, passes through the operator
Can by gating cycle neural network t moment two input feature vector (i.e. utAnd xt) and at last moment (i.e. t-1 moment)
Obtained hiding feature (i.e. ht-1) mapped, to obtain gating cycle neural network in the hiding feature (h of t momentt)。
In formula (1), ztTo update the gating parameter of door, and z can be usedtIt indicates to update activation result corresponding to door,
In other words, the history of last moment can be hidden by feature (i.e. h by the gating parameter of the update doort-1) and current time
Candidate hiding feature (i.e. h' corresponding with the first business datum featuret), first user's Figure Characteristics are corresponding candidate hides spy
Levy (i.e. h "t) be combined, feature is hidden to obtain target shown in fig. 6.In other words, formula (1) obtained target is hidden
Feature can describe the total recall info for finally retaining and remembering in above-mentioned target duration.Total recall info may include
Two parts, a part are (1-zt) and ht-1Obtained time series is by the end of last moment (i.e. t-1 after carrying out step-by-step multiplication
Moment) the historical trace information that retains and remember, another part ztStep-by-step multiplication is carried out between the hiding feature of candidate
The target memory information that obtained current time is newly remembered afterwards.Wherein, symbol " ° " can be used to indicate that the element in matrix
Between can carry out step-by-step multiplication.For example, matrix A=[1,2,3], matrix B=[2,3,4] then carry out matrix A and matrix B
Step-by-step multiplication, it can obtain Matrix C=[1 × 2,2 × 3,3 × 4]=[2,6,12].In other words, which can manage
Solution is multiplied between the element in the same position in different matrixes.
It should be appreciated that in embodiments of the present invention each triggered time stamp can correspond to a service data information and
One customer attribute information.Therefore, it is integrated, can be obtained by the way that this feature is updated each service data information in duration
Sequence is read to above-mentioned business datum, at the same time, each user by the way that this feature to be updated to the same user in duration belongs to
Property information integrated, can also synchronize to obtain and read the corresponding user of sequence with the business datum and draw a portrait evolution sequence, thus
Can when getting first and pulling request, can based on the stamp of each triggered time arrived recorded in above-mentioned data-pushing list,
The business datum is successively read to the business datum feature of each service data information in sequence and is divided with each business datum feature
Not corresponding user's Figure Characteristics are given to the gating cycle neural network so that the gating cycle neural network can be based on it is above-mentioned
Formula (1) obtains stabbing corresponding hiding feature with each triggered time.It is understood that in the gating cycle neural network
Be in corresponding time series, between the corresponding hiding feature of the adjacent triggered time stamp of any two it is associated with each other,
It is updated in duration in this feature, which can will stab obtained hiding spy in a upper triggered time
It levies the history corresponding as stamp of current triggered time and hides feature.
Wherein, in the gating cycle neural network, resetting door 1 and resetting door 2 may be incorporated for control last moment
History hides the hiding feature of candidate how feature flows into current time, and the hiding feature of last moment history may include the time
Whole historical trace information of the sequence by the end of last moment therefore can be with by resetting door shown in fig. 61 and resetting door 2
Abandon unrelated historical trace information.In other words, if it is a feature vector that history, which hides feature, the gate of the resetting door is joined
Number (i.e. it is above-mentioned it is shown in fig. 6 resetting door 1 and reset door 2 gating parameters) may each be one gated composed by 0 to 1 to
Amount, the gating parameter of the resetting door can be used for measuring this feature and update the size that the gate in duration is opened.For example, with resetting
The gating parameter g of door 2tFor, if updating in duration in feature, if the history at the t-1 moment hides feature, (history hides spy
May include: age characteristics, sex character, regional feature etc. in sign) in there are in the gate vector corresponding to age characteristics
Gate value is 0, then then can choose the age characteristics for forgeing the t-1 moment by the resetting door 2 at current time, and retaining should
History hides other features in feature.Similarly, if the history at the t-1 moment hides the gate for resetting door 2 corresponding to feature
Parameter is 0, hides feature then then can choose by the resetting door 2 at current time and forget the history of last moment.Wherein, should
Following formula (2) can be used to be calculated in these three gating parameters in gating cycle neural network:
[zt,rt,gt]=σ ([Wz,Wr,Wg]xt+[Vz,Vr,Vg]ut+[Uz,Ur,Ug]ht-1+[bz,br,bg]) formula (2)
In formula (2), rtFor the gating parameter for resetting door 1, gtFor the gating parameter for resetting door 2, ztFor the door for updating door
Control parameter, Wz, Wr, WgIt is the input feature vector x of the gating cycle neural networktCorresponding to (i.e. the first business datum feature)
The first weight parameter in weight matrix;In formula (2), Vz, Vr, VgIt is the input feature vector of the gating cycle neural network
utThe second weight parameter in weight matrix corresponding to (i.e. first user's Figure Characteristics);In formula (2), Uz, Ur, UgFor
The input feature vector h of the gating cycle neural networkt-1Third weight ginseng in weight matrix corresponding to (i.e. history hides feature)
Number;In formula (2), bz, br, bgFor the straggling parameter of the gating cycle neural network.Wherein, it should be understood that the gating cycle
Above-mentioned four partial information can be added by the update door in neural network, and be put into activation primitive (i.e. sigma function), so as to
The corresponding activation result of each gating parameter to be compressed between 0 to 1.In other words, it in the target network model, can obtain
Take corresponding first weight parameter of the first business datum feature, corresponding second weight parameter of the first user data feature,
The history hides the straggling parameter of the corresponding third weight parameter of feature and the target network model, and can be according to
Above-mentioned formula (2) is by the first business datum feature, first weight parameter, the first user Figure Characteristics, described
Second weight parameter, the history hide feature, the third weight parameter and the straggling parameter as the target network
The corresponding input parameter of activation primitive in network model, so as to export the corresponding gating parameter of the update door, described
The corresponding gating parameter of one resetting door and the corresponding gating parameter of the second resetting door.
Wherein, as shown in fig. 6,1 (i.e. h' of candidate hidden state in the gating cycle neural networkt) and candidate hiding shape
2 (i.e. h " of statet) calculation formula following formula (3) and formula (4) can be used to indicate:
h't=tanh (Wh'xt+Uh'(rt°ht-1)+bh') formula (3)
h”t=tanh (Wh”ut+Uh”(gt°ht-1)+bh”) formula (4)
Symbol " ° " in formula (3) can be used to indicate that gating parameter rtFeature h is hidden with historyt-1Between can root
Operation is carried out according to the step-by-step multiplication of matrix.Similarly, the symbol " ° " in formula (4) can be used to indicate that gating parameter gtWith go through
History hides feature ht-1Between can according to the step-by-step multiplication of matrix carry out operation.In formula (3), Wh', Uh'And bh'It is this
For obtaining candidate 1 (i.e. h' of hidden state in gating cycle neural networkt) corresponding candidate parameter, similarly, in formula (4)
In, Wh”, Uh”And bh”It is in the gating cycle neural network for obtaining candidate 2 (i.e. h " of hidden statet) corresponding to candidate
Parameter.In other words, feature and first business are hidden based on the corresponding gating parameter of the first resetting door, the history
Data characteristics can determine the corresponding candidate hiding feature of the first business datum feature in the target network model;
In addition, hiding feature and first user portrait spy based on the corresponding gating parameter of the second resetting door, the history
Sign can determine the corresponding candidate hiding feature of the first user Figure Characteristics in the target network model.
It should be appreciated that as shown in Figure 6 above, input feature vector xtIt can be understood as business datum constructed by destination server
Feature vector is (for example, X={ x1, x2... xk, xt-1, xt) in t-th of component, wherein the value of k and t is just greater than 0
Integer.It can be by vector { x in the business datum feature vector, X1, x2... xkIt is referred to as above-mentioned first business trigger list pair
The historical data feature vector answered;It can be by vector xt-1Referred to as second business number of the target network model in the historical juncture
It, can also be by vector x according to featuretReferred to as first business datum feature of the target network model in object time.It should manage
Service data information each in the above-mentioned second trigger list corresponding triggered time can be stabbed institute's structure in the embodiment of the present invention by solution
At sequence be referred to as time series, when so as to by stamp of each triggered time being referred to as one in the time series
Quarter or a time step.In consideration of it, if it is to arrive recorded in above-mentioned table 1 that this feature, which updates multiple first moment in duration,
Three triggered times stab (i.e. A41Moment, A42Moment, A43Moment), then the destination server can be in these three triggered times
The maximum triggered time is stabbed into corresponding first moment in stamp and is referred to as object time, and can by before object time first when
It carves (or time step before current time step) and is referred to as the historical juncture, and can be (i.e. shown in fig. 6 by the target network model
Gating cycle neural network) the maximum triggered time in duration is updated in a upper feature stabs corresponding second moment (i.e. aforementioned industry
Triggered time stamp corresponding to k-th of component of business data characteristics vector X) it is referred to as initial time.In other words, the present invention is implemented
This feature can be updated the stamp of triggered time corresponding to the previous time step of duration and be referred to as initial time by example, and can be with broad sense
On think in the time series corresponding to above-mentioned target duration, the initial time be the historical juncture previous moment, i.e.,
The hiding feature exported when getting second and pulling request, at the time of can be by corresponding to aforementioned k-th of component is referred to as
Initially hiding feature corresponding with this feature update duration.
For ease of understanding, further, Fig. 7 is referred to, is that a kind of acquisition target provided in an embodiment of the present invention hides spy
The schematic diagram of a scenario of sign.Multiple first in duration are updated in feature comprising the target user in time series as shown in Figure 7
Moment corresponding user's Figure Characteristics and corresponding business datum feature of multiple first moment, pass through the target network
Model can the User Status feature (i.e. user's Figure Characteristics and business datum feature) to each first moment model, from
And target shown in Fig. 7 can be exported and hide feature.Wherein, for ease of understanding, the embodiment of the present invention updates duration with this feature
Three triggered times that interior multiple first moment are recorded by above-mentioned table 1 stab (A i.e. shown in Fig. 741Moment, A42Moment,
A43Moment) for, to describe the incidence relation between upper feature update duration and this feature update duration.On as shown in fig. 7,
Multiple second moment that one feature updates in duration may include A shown in Fig. 731Moment, A32Moment.In the time series,
A32Moment is at the time of stabbing corresponding in the maximum triggered time that feature updates in duration on this, so as to by A shown in Fig. 732
Moment is referred to as initial time, and the target network model can be updated the obtained hiding feature of duration in a upper feature
(hiding feature 5 i.e. shown in Fig. 7) updates the A in duration as this feature41Moment (A41Moment can be referred to as shown in Fig. 7
Time series in the 1st time step) it is initial hide feature, and can be by the second user Figure Characteristics of the 1st time step
(user's Figure Characteristics 1 i.e. shown in Fig. 7) and the second business datum feature (business datum feature 1 shown in Fig. 7) are initial with this
Hiding feature is merged, to export the A41Moment corresponding hiding feature (hiding feature 1 i.e. shown in Fig. 7).Wherein, the mesh
The concrete mode that mark server merges each input feature vector is referred to above-mentioned formula (1), formula (2), formula (3) and public affairs
Formula (4), so as to obtain the target network model in the hiding feature 1 of the 1st time step.Further, the destination server
The target network model can be used as in the 2nd time step (i.e. Fig. 7 in the hiding feature 1 of the 1st time step on the target network model
Shown in A42Moment) input feature vector, and can be using the corresponding business datum feature of the 2nd time step as the second new business
Data characteristics (business datum feature 2 i.e. shown in Fig. 7), at the same time, which can also be by the 2nd time step pair
The user's Figure Characteristics answered can will be obtained as new second user Figure Characteristics (user's Figure Characteristics 2 i.e. shown in Fig. 7)
The the second new business datum feature got and new second user Figure Characteristics are used as the target network model at the 2nd together
Spacer step (i.e. A42Moment) input feature vector, so as to so that the target network model pass through above-mentioned formula (1), formula (2), formula
(3) and formula (4) obtains the target network model in A42The output feature (hiding feature 2 i.e. shown in Fig. 7) at moment, it can
The target network model is obtained in the hiding feature 2 of the 2nd time step.
It should be appreciated that the embodiment of the present invention can be by the target network model in the obtained hiding feature 1 of the 1st time step
It may be collectively referred to as history with the obtained hiding feature 2 of the 2nd time step and hide feature, and can be by A shown in Fig. 741Moment and
A42Moment is referred to as the historical juncture.Further, which can be by the target network model in the hidden of the 2nd time step
Feature 2 is hidden as the target network model in the 3rd time step (i.e. A43Moment) input feature vector, and can be by the 3rd time step
Corresponding business datum feature is as the first business datum feature (business datum feature 3 i.e. shown in Fig. 7), at the same time, should
Destination server can also be using the corresponding user's Figure Characteristics of the 3rd time step as first user's Figure Characteristics (i.e. shown in Fig. 7
User's Figure Characteristics 3), and the first business datum feature that can be will acquire and first user's Figure Characteristics together as should
Target network model is in the 3rd time step (i.e. A43Moment) input feature vector, so as to so that the target network model by above-mentioned
Formula (1), formula (2), formula (3) and formula (4) obtain the target network model in A43(output is special for the output feature at moment
Sign can be hiding feature 3 shown in Fig. 7), it can the target network model is obtained in the hiding feature 3 of the 3rd time step, and
The target that the hiding feature 3 of 3rd time step can be referred to as to the target network model hides feature.At this point, the target is hidden
Hiding feature can be referred to as the hiding feature of new first (or second hiding feature).It should be appreciated that the hiding feature of the target can
To pull the user for requesting corresponding target user portrait information with described first for characterizing.It can be seen that by will be continuous
User's Figure Characteristics and business datum feature in period model, available more complete user's portrait information,
Information is drawn a portrait when being subsequently used for carrying out data recommendation so as to ensure that the target hides the user that feature is characterized, Neng Gouyou
Effect ground ensures that recommended data result can be bonded the hobby of the target user, it can improves the accuracy of data recommendation.
In embodiments of the present invention, response first pulls request for target application, obtains this and first pulls and request pair
The feature answered updates duration, this feature when updating a length of this first pull request and upper one and pull duration between request, the spy
Sign, which updates, may include multiple first moment in duration, the first moment of each of the multiple first moment is corresponding corresponding
Business datum feature and user's Figure Characteristics;So can be by target network model obtained by a upper feature updates in duration
Hiding feature be determined as initially hiding feature, so as to will initially hide feature, described more using the target network model
A first moment corresponding first user Figure Characteristics and corresponding first business number of the multiple first moment
It is merged according to feature, is exported so that the target network model can update in duration in this feature for describing target user's
The target of user's portrait information hides feature.Wherein, it can establish in continuous time period and (i.e. should by the target network model
Feature update duration and a upper feature update duration) user's Figure Characteristics and business datum feature between incidence relation, i.e.,
The target of target network model output, which hides feature, can combine the user under the different moments in continuous time period
Figure Characteristics and business datum feature are more bonded target use so as to accurately depict in business datum recommender system
User's portrait information of family hobby, and then can effectively ensure that the accuracy and completeness of user's representation data.
Further, Fig. 8 is referred to, is the stream of another user's representation data processing method provided in an embodiment of the present invention
Journey schematic diagram.As shown in figure 8, the method may include:
Step S201, response pull request for the first of target application, obtain described first and pull the corresponding spy of request
Sign updates duration.
Wherein, it includes multiple first moment in duration that the feature, which updates,;It is by described first that the feature, which updates duration,
Request corresponding first request time stamp is pulled to pull determined by corresponding second request time of request stabs with second;Described
Two, which pull request, pulls upper the one of request for described first and pulls request, and described second pulls request for acquisition and target user
Multiple history recommending data information with incidence relation.
Step S202, the feature update duration in, by the multiple history recommending data information with the target
History recommending data information between user there are data-triggered relationship is determined as target histories recommending data information.
Specifically, destination server can be obtained in data-pushing list updates be pushed to target in duration for this feature
Multiple history recommending data information of user, so as to from multiple history recommending data information will with the target user it
Between there are the history recommending data information of data-triggered relationship to be determined as target histories recommending data information, it can will carry
The history recommending data information for clicking label is referred to as target histories recommending data information.
Step S203 is based on each target histories recommending data information and history service data information, constructs the mesh
Mark the corresponding business datum feature vector of network model;
Specifically, each target histories recommending data information can be added to the first business trigger column by destination server
Table, and the first business trigger list of each target histories recommending data information will be carried as the second business trigger column
Table;It is gone through comprising existing between the target user in the first business trigger list and the second business trigger list
The history service data information of history triggering relationship;Further, in the second business trigger list, the feature is updated
The first moment, corresponding target histories recommending data information was determined as target service data information each of in duration;Further
Ground extracts the semantic feature of the target service data information, and using the semantic feature extracted as the target
The corresponding target service data characteristics of service data information;Further, acquisition the first business trigger list is corresponding goes through
History data characteristics vector, and the target service data characteristics and the historical data feature vector are spliced, and will spell
Historical data feature vector after connecing is determined as the corresponding business datum feature vector of the target network model.Wherein, described
The corresponding business datum feature of multiple first moment in duration is updated comprising the feature in business datum feature vector;
The corresponding target histories recommending data information of one business datum feature.
Step S204, by the corresponding target user's attribute information of each target histories recommending data information and
The corresponding historical user's attribute information of the history service data information constructs the corresponding user's portrait of the target network model
Feature vector;
Specifically, destination server can be in the second business trigger list, by the mesh at each first moment
Customer attribute information corresponding to mark history recommending data information is determined as target user's attribute information;It further, will be described
User property feature in target user's attribute information is combined with the attention model feature in the target network model,
And the user property feature after combination is determined as the corresponding target user's Figure Characteristics of target user's attribute information;Into one
Step ground, obtains historical user's feature vector associated with the first business trigger list, and the target user is drawn a portrait
Feature is spliced with historical user's feature vector, and spliced historical user's feature vector is determined as the target
The corresponding user's Figure Characteristics vector of network model.Wherein, when in user's Figure Characteristics vector comprising the multiple first
Carve corresponding user's Figure Characteristics.
For ease of understanding, further, Fig. 9 is referred to, is that a kind of building user portrait provided in an embodiment of the present invention is special
Levy the schematic diagram of a scenario of vector.As shown in figure 9, destination server can by gender shown in Fig. 9, the age, region, label ...,
The attribute informations such as classification are referred to as to be input to the customer attribute information of vector embeding layer shown in Fig. 9.Wherein, the present invention is implemented
Example can claim the target user in the customer attribute information at each first moment (for example, t moment, t is the positive integer greater than 0)
Be target user's attribute information.So when that target user's attribute information is input to vector shown in Fig. 9 is embedding for destination server
After entering layer, the vector embeding layer can be made to export user property feature corresponding with each attribute information.Wherein, user property feature
It may include: (sex character, male);(age characteristics, 50);(regional feature, Jiangsu);(label characteristics, sketch, tomato are prevented
Cancer);(classification feature, health-preserving, cuisines).It is understood that the feature in the embodiment of the present invention updates in duration, it can be with
Corresponding customer attribute information of each first moment (i.e. each triggered time stamp) is input to vector embeding layer shown in Fig. 9
It afterwards, can one user property feature of corresponding output.Further, which can be in t moment by target user
User property feature in attribute information be given to it is shown in Fig. 9 from attention layer, so as to belong to from above-mentioned multiple users
The corresponding target user's Figure Characteristics of target user's attribute information are determined in property feature.For example, the target user is in t moment
Target user's Figure Characteristics can be with are as follows: (sex character, male);(age characteristics, 50);(regional feature, Jiangsu).It should manage
Features described above can be updated different moments obtained user's Figure Characteristics in duration and be referred to as mesh by solution, the embodiment of the present invention
User's Figure Characteristics are marked, and can further be spelled target user's Figure Characteristics and historical user's feature vector
It connects, and spliced historical user's feature vector is determined as the corresponding user's Figure Characteristics vector of the target network model,
To construct user's Figure Characteristics vector shown in Fig. 9.
Wherein, from attention layer, user property feature can be denoted as F=[f1, f2..., fn], at this point it is possible to use this
A little attributive character indicate a series of user characteristics, at this point, should be from the attention model of attention layer from attention mechanism
It can indicate are as follows:
In formula (5), Q, K, V are the model parameter of attention model, and have Q=K=V=X.In formula (5), d
For the length of above-mentioned user property feature vector F.In addition, the embodiment of the present invention using bull from attention mechanism, so as to
Above-mentioned each user property feature is combined, and does not have ordinal characteristics between each user property feature, so as to obtain
Meet user's Figure Characteristics u of user's attention ratet.A head can indicate in multiple heads employed in the attention mechanism
Are as follows:
headi=Attention (QWi Q,KWi K,VWi V) formula (6)
In formula (6), Wi Q, Wi K, Wi VFor the weight coefficient in weight matrix W corresponding to each model parameter;Cause
This, updates in duration in features described above, multiple heads can be spliced, then be multiplied with above-mentioned matrix W, it can obtains and is somebody's turn to do
Feature, which updates the corresponding user of each moment in duration, indicates vector
Wherein, user indicates that vector can be described as:
U1=[head1,head2,head3,...,headn] W ° of formula (7)
It should be appreciated that the user in formula (7) indicates vector U1It can be in constructed user's Figure Characteristics vector
Component of a vector corresponding to each moment in duration is updated with feature, when each component of a vector can be understood as features described above update
The corresponding user's Figure Characteristics of target user's attribute information in length.The embodiment of the present invention can further by the user indicate to
Historical user's feature vector described in (i.e. U1) and history is measured ((for example, user's portrait at the first moment of U2, each of the U2 is special
Sign similarly is using above-mentioned attention model and obtain)) spliced, and then can by spliced historical user's feature to
Amount is determined as the corresponding user's Figure Characteristics vector of the target network model.
Target network model is updated obtained hiding feature in duration in a upper feature and is determined as initially by step S205
Hide feature;
Wherein, the initial hiding feature is initial time of the target network model in upper feature update duration
Obtained hiding feature;The stamp of maximum triggered time that the initial time can update in duration for a upper feature corresponds to
The second moment;A length of above-mentioned second pulls request and second pulls upper the one of request with this and pull and ask when a upper feature updates
The target signature between (the i.e. above-mentioned 4th pulls request) is asked to update duration.
It is special to obtain corresponding business datum of the multiple first moment from business data characteristics vector by step S206
Sign, and corresponding user's Figure Characteristics of the multiple first moment are obtained from user's Figure Characteristics vector;
Wherein, the multiple first moment includes historical juncture and object time, and the multiple first moment respectively corresponds
Business datum feature in include the historical juncture corresponding second business datum feature, the object time corresponding first
Business datum feature;Include the historical juncture corresponding the in the multiple first moment corresponding user's Figure Characteristics
The corresponding first user Figure Characteristics of two user's Figure Characteristics, the object time;The historical juncture is the object time
Last moment.
Step S207, according to the initially hiding feature, corresponding business datum feature of the multiple first moment,
And the multiple first moment corresponding user's Figure Characteristics, determine that the target network model is updated in the feature
Target in duration hides feature.
Specifically, destination server can be by the target network model by second user Figure Characteristics, described
Second business datum feature and the initially hiding feature are merged, and using fused initially hiding feature as described in
History of the target network model in the historical juncture hides feature;It further, will be described by the target network model
First user's Figure Characteristics, the first business datum feature and the history are hidden feature and are merged, and will be after fusion
History hide feature as the target network model the object time target hide feature.Wherein, target is obtained
The specific acquisition process for hiding feature may refer to the description for hiding feature in embodiment corresponding to above-mentioned Fig. 7 to target, here
It will not continue to repeat.
Step S208 hides the user portrait information that feature is characterized by the target and obtains from service database
Take multiple first candidate data information;
Step S209, obtains the target user corresponding preliminary screening rule, and from the multiple first candidate data
The first candidate data information that information sifting is consistent with the preliminary screening rule is as the second candidate data information;
Step S210, selection is hidden feature with the target and is characterized from all second candidate data information filtered out
User's Figure Characteristics vector, the second candidate data information for being consistent of the business datum feature vector is as number to be recommended
It is believed that breath.
Selected data information to be recommended is handed down to the target as target recommending data information and used by step S211
Family.
Wherein, the target recommending data information may include multiple sub- recommending data information, when destination server is by institute
Multiple sub- recommending data information are stated to be pushed to after the target user, it can be according in the multiple sub- recommending data information
The click probability of every sub- recommending data information generates the corresponding predicted vector of the first clicking rate prediction model, wherein institute
Stating predicted vector includes target prediction vector;The target prediction vector pulls request described first for describing target user
Subsequent time prediction probability that the first sub- recommending data information in the multiple sub- recommending data information is clicked;Into
One step, it is sub for second in the multiple sub- recommending data information in the subsequent time when the target user is recorded
When click relationship (the i.e. new data-triggered relationship) of recommending data information, described first is generated based on the click relationship and is clicked
The corresponding click vector of rate prediction model;The click vector includes the target point amount of hitting to;Wherein, the target point amount of hitting to is used
The second sub- recommending data in the multiple sub- recommending data information is believed in the subsequent time in describing the target user
Cease the actual probabilities clicked;Further, described in the target point amount of hitting to according to the target prediction vector sum determines
The loss function of first clicking rate prediction model, and the loss function is optimized by gradient decreasing function, and will be excellent
Loss function after change is determined as objective function, and the corresponding first clicking rate prediction model of the objective function is determined as
Two clicking rate prediction models.It is understood that the first clicking rate prediction model and the second clicking rate in the embodiment of the present invention
Prediction model can be the clicking rate prediction model in embodiment corresponding to above-mentioned Fig. 2 b.
Wherein, before the target recommending data information filtered out is handed down to target user by destination server, target clothes
Business device can also predict each of selected data information to be recommended data to be recommended by the first clicking rate prediction model
The corresponding click probability of information, and the corresponding click probability of each data information to be recommended is ranked up, and from sequence
Target recommending data information is obtained in data information to be recommended afterwards, so as to more be bonded the target user for selected
The target recommending data information recommendation of hobby gives the target user.
For ease of understanding, further, referring to Figure 10, it is that a kind of acquisition target provided in an embodiment of the present invention recommends number
It is believed that the schematic diagram of a scenario of breath.Wherein, for ease of understanding, the embodiment of the present invention is using the service data information as article information
Example, when target user executes trigger action to the operating area where target application in the first display page shown in Fig. 10
When, the corresponding article information acquisition request of the trigger action, this article can be generated in target terminal user shown in Fig. 10
Information acquisition request can pull request for shown in Fig. 10 first, and further, which can be to Figure 10 institute
The destination server shown sends this and first pulls request.It should be appreciated that the destination server first pulls request getting this
Later, can based on this first pull request corresponding first pull timestamp and second pull request corresponding second pull when
Between stab, determine with target user shown in Fig. 10 have incidence relation target network model feature update duration, thus
It can be updated in duration in this feature, the customer attribute information of the target user is modeled, is existed with obtaining the target user
User's Figure Characteristics vector in above-mentioned target duration.In addition, the destination server can also update in duration in this feature, it is right
The article information that the target user read models, so as to obtain the target user in above-mentioned target duration
Business datum feature vector.
Wherein it is possible to understand, updated in duration in this feature, which clicks all articles read
The sequence information constituted can be referred to as above-mentioned business datum and read sequence, and can synchronize to obtain business datum reading sequence
Corresponding user's portrait evolution sequence is arranged, and then business datum can be read to business corresponding to each article information in sequence
User's Figure Characteristics corresponding to each customer attribute information in data characteristics and user's portrait evolution sequence, as the target network
The input feature vector of network model, it is defeated corresponding to variant triggered time stamp so as to be obtained based on the target network model construction
Enter the incidence relation between feature.It should be appreciated that the incidence relation between these features can be existed with the target network model
Feature updates the target exported in duration and hides feature to describe.In consideration of it, as shown in Figure 10, which is obtaining
After the target hides feature, the association between the input feature vector at variant moment described in feature can be hidden based on target
Relationship (i.e. the target hides user's portrait information that feature is currently characterized), filters out from service database shown in Fig. 10
There is the data information to be recommended of incidence relation with the target user, wherein the data information to be recommended may include Figure 10
Shown in article information 10, article information 20, article information 30, article information 40, article information 50, article information 60.
Wherein, which, which can be used clicking rate prediction model and estimate target user, filters out above-mentioned Figure 10
The probability clicked of data information to be recommended, and can be further corresponding to these data informations to be recommended estimated
Click probability be normalized, so as to be ranked up to the click probability after normalization, in order to can be from row
Multiple data informations to be recommended that sequence forward (i.e. maximum probability) is selected in these click probability after sequence are pushed away as target
Data information is recommended, so as to which the target recommending data information is handed down to target terminal user shown in Fig. 10.For example, target
Server can predict the corresponding click probability of 6 article informations shown in Fig. 10 according to the clicking rate prediction model, and can
It is normalized with the click probability further to this 6 article informations, obtains this 6 business attribute informations and respectively correspond
New click probability, so as to select after the new click probability to this 6 service data informations is ranked up
The multiple article informations (for example, two article informations that can will come the first two) for providing maximum probability recommend number as target
It is believed that breath recommends target user.
Wherein, clicking rate prediction model can be the first clicking rate prediction model, the first clicking rate prediction model it is pre-
Estimating function can be softmax function, which can be used for hiding feature according to target sieves from service database
The data information to be recommended for meeting user interest is selected, so as to further predict target user at lower a moment to being screened
The probability that each of data information to be recommended out data information to be recommended is clicked.In consideration of it, first clicking rate is pre-
Estimating model can indicate are as follows:
P=softmax (ht) formula (8)
In other words, available each to the corresponding prediction probability of data information to be recommended (on i.e. by the formula (8)
State the click probability estimated).For ease of understanding, the data to be recommended that the embodiment of the present invention is filtered out with destination server
Information is above-mentioned article information 10, article information 20, article information 30, article information 40, article information 50, article information 60 are
Example, to illustrate the another way for obtaining target recommending data information.Further, table 2 is referred to, is that the embodiment of the present invention mentions
A kind of distribution situation table of the click probability of the data information to be recommended filtered out supplied.
As described in Table 2, the click probability of article information 10, article information 20, article information 50 and article information 60
It is all larger than the click probability threshold value of the first clicking rate prediction model, therefore, which can be by article information 10, text
Chapter information 20, article information 50 and article information 60 are referred to as target recommending data information, so as to which the target is recommended number
It is believed that breath is handed down to target terminal user shown in Fig. 10, so that the target terminal user can be in the second display shown in Fig. 10
Aforementioned document information 10, article information 20, article information 50 and article information 60 are shown on the page.
Table 2
It should be appreciated that further, target user can also recommend the target shown in the target terminal user
Data information executes trigger action, so as to generate new data-triggered instruction based on the trigger action, and by the new number
It is given to the destination server according to triggering command, so that destination server can be further to the target in above-mentioned data-pushing list
The corresponding state tag of recommending data information is updated, so as to obtain new data-pushing list.Therefore, when the target
When server gets third and pulls request, first it can pull that request corresponding first pulls timestamp and third is drawn based on this
It takes the corresponding third of request to pull timestamp, determines that (it is corresponding next that i.e. above-mentioned third pulls request to new feature update duration
Feature updates duration), and the feature feature new as this can be hidden more for target associated by duration is updated with features described above
The initial of new historical juncture in new duration hides feature, so as to repeat above-mentioned steps S101- step S104, with
It obtains new target and hides feature, so as to recommend new article information to be given to the target user, to ensure data recommendation
Accuracy.
Wherein it is possible to understand, since the function softmax function of estimating in clicking rate prediction model is for predicting
The probability that target user clicks each sub- recommending data information issued at lower a moment, this probability can be understood as claiming
To click probability or prediction probability.Therefore, the embodiment of the present invention can take the mode of supervision to estimate the clicking rate
The function of estimating of model exercises supervision, for example, when the clicking rate prediction model predicts user in subsequent time to article information
xkThe prediction probability clicked is maximum, however is carved into for the moment when coming instantly, and user does not simultaneously click on this and has maximum predicted probability
Article information xk, but another article information of clicking rate xi, therefore, when the article serial number k in above-mentioned predicted vector is not equal to
When clicking the article serial number i in vector, i.e., when prediction result is not consistent with actual result, then need to the clicking rate prediction model
Currently used function of estimating is modified, and the method so as to use gradient to decline estimates function to currently used
Optimize, so as to when the model parameter of the clicking rate prediction model reaches convergence, determine obtain it is new estimate function,
So as to which the new function of estimating is referred to as objective function.
Wherein, for ease of understanding, the embodiment of the present invention can by taking 6 data informations to be recommended shown in above-mentioned table 2 as an example,
Pass through estimating function and can predict to obtain each data to be recommended letter in this 6 data informations to be recommended in above-mentioned formula (5)
Corresponding click probability is ceased, that is, has Y={ y1, y2, y3, y4, y5, y6 }.
Wherein, y1 is the click probability (i.e. 0.15) of article information 10 shown in above-mentioned table 2, and y2 is shown in above-mentioned table 2
The click probability (i.e. 0.25) of article information 20, y3 are the click probability (i.e. 0.1) of article information 30 shown in above-mentioned table 2, y4
For the click probability (i.e. 0.1) of article information 40 shown in above-mentioned table 2, y5 is the click of article information 50 shown in above-mentioned table 2
Probability (i.e. 0.25), y6 are the click probability (i.e. 0.15) of article information 60 shown in above-mentioned table 2.Since first clicking rate is pre-
The click probability threshold value for estimating model is 0.1, and therefore, the embodiment of the present invention can be by the click of 6 article informations in above-mentioned table 1
Probability is ranked up, and by the article information 20 of sequence symbol, article information 50, article information 10 and article information 60 are referred to as mesh
Recommending data information is marked, in other words, which may include 4 sub- recommending data information, work as destination service
When above-mentioned 4 sub- recommending data information are handed down to target user by device, it can be recommended with the synchronous recording target user and this 4 sons
Data-pushing relationship between data information, so as to recommend the son between the target user there are data-pushing relationship
Data information is referred to as the first sub- recommending data information, so as to according to the point of this 4 sub- recommending data information pushed
It hits probability and generates the corresponding predicted vector of the first clicking rate prediction model, for example, P={ y2, y5, y1, y6 }=0.25,
0.25,0.15,0.15 }={ p1, p2, p3, p4 }.Wherein, p1 is the corresponding prediction probability of article vector 20, and p2 is article vector
50 corresponding prediction probabilities, p3 are the corresponding prediction probability of article vector 10, and p4 is the corresponding prediction probability of article vector 60.Its
In, p1, p2, p3, p4 may be collectively referred to as target prediction vector.It is understood that can be from filtering out in the embodiment of the present invention
N number of (6 i.e. above-mentioned) data information to be recommended in obtain k (4 i.e. above-mentioned) sub- recommending data information and be handed down to target use
Family.
It should be appreciated that after above-mentioned 4 sub- recommending data information are handed down to target user by the destination server, Ke Yijin
One step statistics updates the target user in duration and the click relationship between this 4 sub- recommending data information in new feature, from
And the sub- recommending data information between the target user with click relationship can be referred to as the second sub- recommending data information,
And then the corresponding click vector of the first clicking rate prediction model can be generated based on the click relationship.For example, the target
When user is directed to the execution clicking operation of article information 10 in the 4 sub- recommending data information issued, which can be with table
It is shown as: R={ r1, r2, r3, r4 }={ 0,0,1,0 }., wherein r1 is the corresponding actual probabilities of article vector 20, and r2 is article
The corresponding actual probabilities of vector 50, r3 are the corresponding actual probabilities of article vector 10, and r4 is that the corresponding reality of article vector 60 is general
Rate.Wherein, r1, r2, r3, r4 may be collectively referred to as the target point amount of hitting to.
It is understood that at this point, destination server can be based on the target point amount of hitting to described in the target prediction vector sum
Determine the loss function of the first clicking rate prediction model, the calculation formula of the loss function can be with are as follows:
Loss (P, R)=∑0≤i≤NRi*log (pi) formula (9)
In formula (9), P is above-mentioned predicted vector, and R is above-mentioned click vector, and i is the element number in each vector, and N is
It is pushed to the number of the sub- recommending data information of the target user.Pi is the target prediction vector in predicted vector, and the target is pre-
Direction finding amount be used for describe get described first pull request after subsequent time in the multiple sub- recommending data information
The prediction probability clicked of the first sub- recommending data information.Similarly, ri is the target point amount of hitting to clicked in vector, the mesh
The punctuate amount of hitting to is for describing the target user in the subsequent time to second in the multiple sub- recommending data information
The actual probabilities that sub- recommending data information is clicked, the value of the actual probabilities are 0 (not clicking on) or 1 (click).
Wherein, after destination server can pull request getting above-mentioned first, it is based on business datum recommender system
In recall logic, primary election logic and smart sequence logic screen the mesh characterized with the hiding feature of the target from service database
The data information to be recommended that user's portrait information of mark user is consistent, and the data information to be recommended filtered out is ranked up,
So as to which above-mentioned target recommending data information is handed down to the target user, so that the corresponding target user of the target user is whole
Received target recommending data information can be shown by end.
For ease of understanding, further, 11 are referred to, is a kind of business datum recommender system provided in an embodiment of the present invention
Schematic diagram.As shown in figure 11, it may include the data system of information for storing data in the business datum recommender system, it should
It may include service database 10d shown in Figure 11 and user's portrait library (i.e. portrait of user shown in Figure 11 in data system
10c), user portrait library can be user terminal each in embodiment corresponding to above-mentioned Fig. 1 (for example, user terminal 3000a, is used
Family terminal 3000b ..., user terminal 3000n) corresponding user user draw a portrait evolution information (i.e. the user draw a portrait
Library can be used for storing user's portrait evolution information that customer attribute information of the same user in different moments is constituted).Work as figure
Destination server shown in 11 gets transmitted by target terminal user (for example, user terminal 3000a) first and pulls request
When, it can be drawn a portrait based on the method 1 (for example, user's portrait recalls method) of recalling recalled shown in Figure 11 in logic from user
User's portrait evolution information of target user corresponding to user terminal 3000a is obtained in 10c, so as to according to the user
Portrait evolution information obtain the target user specific portrait information (for example, the age: 30, gender: female, region: Shenzhen, Wuhan,
Shanghai), in order to which the business datum being consistent with the specific portrait information can be obtained from service database 10d shown in Figure 11
Information, in other words, destination server can recall method based on user portrait and pull multiple first from the service database
As recalling shown in Figure 11, logic 10a is corresponding to recall result to candidate data information.
Further, which largely can recall result and be given to primary election shown in Figure 11 from what will be pulled
Logic 10b, so as to according in primary election logic 10b preliminary screening rule (such as: customer documentation correlation (the document
Correlation be the correlation between different document is carried out by content tuning algorithm it is obtained after cooperated computing), it is effective,
The screening rules such as regional and diversity) preliminary screening is carried out to above-mentioned multiple first candidate data information, it is first with this to obtain
The first candidate data information that step screening rule is consistent is as the second candidate data information.Wherein, screening rule shown in Figure 11
1,., screening rule n be screening rule in aforementioned preliminary screening rule, can be from largely calling together by primary election logic 10b
It returns in result and filters out the personalized recommendation information to match with the target user, to efficiently reduce subsequent fine sequence logic
The calculation scale of 10c.Further, which can be by essence sequence logic 10c shown in Figure 11 from filtering out
The user's Figure Characteristics vector characterized with the hiding feature of the target, described is selected in all second candidate data information
The second candidate data information that business datum feature vector is consistent is as data information to be recommended.So as to by shown in Figure 11
Prediction model (i.e. above-mentioned clicking rate prediction model) the click probability of the data information to be recommended filtered out is ranked up,
So as to select the data to be recommended clicked probability and be greater than click probability threshold value from the data information to be recommended after sequence
Information is as target recommending data information, in order to the target recommending data information is pushed to where the target user
User terminal is shown.The use of different user it should be appreciated that the embodiment of the present invention can also draw a portrait to above-mentioned user in library
Similitude between the portrait information of family carries out cooperated computing, so as to which the target recommending data information to be pushed to and the target
User has similar features other users.
In embodiments of the present invention, response first pulls request for target application, obtains this and first pulls and request pair
The feature answered updates duration, this feature when updating a length of this first pull request and upper one and pull duration between request, the spy
Sign, which updates, may include multiple first moment in duration, the first moment of each of the multiple first moment is corresponding corresponding
Business datum feature and user's Figure Characteristics;So can be by target network model obtained by a upper feature updates in duration
Hiding feature be determined as initially hiding feature, so as to will initially hide feature, described more using the target network model
A first moment corresponding first user Figure Characteristics and corresponding first business number of the multiple first moment
It is merged according to feature, is exported so that the target network model can update in duration in this feature for describing target user's
The target of user's portrait information hides feature.Wherein, it can establish in continuous time period and (i.e. should by the target network model
Feature update duration and a upper feature update duration) user's Figure Characteristics and business datum feature between incidence relation, i.e.,
The target of target network model output, which hides feature, can combine the user under the different moments in continuous time period
Figure Characteristics and business datum feature are more bonded target use so as to accurately depict in business datum recommender system
User's portrait information of family hobby, and then can effectively ensure that the accuracy and completeness of user's representation data.
Further, referring to Figure 12, it is a kind of knot of user's representation data processing unit provided in an embodiment of the present invention
Structure schematic diagram.User's representation data processing unit 1 can be applied to above-mentioned destination server, which can be
Service server 2000 in embodiment corresponding to above-mentioned Fig. 1.Further, which can wrap
Include: duration obtains module 10, and first obtains module 20, and second obtains module 30, target signature determining module 40;Further,
User's representation data processing unit 1 can also include: triggering relationship determination module 50, and traffic vector constructs module 60, use
Family vector constructs module 70, data acquisition module 80, the first screening module 90, the second screening module 100 and data distributing module
110;
Duration obtains module 10, pulls request for the first of target application for responding, acquisition described first, which pulls, asks
Corresponding feature is asked to update duration;It includes multiple first moment in duration that the feature, which updates,;
First obtains module 20, for target network model to be updated obtained hiding feature in duration in a upper feature
It is determined as initially hiding feature;
Second obtains module 30, corresponding for obtaining the multiple first moment from business data characteristics vector
Business datum feature, and corresponding user's portrait spy of the multiple first moment is obtained from user's Figure Characteristics vector
Sign.
Wherein, the multiple first moment includes historical juncture and object time, and the multiple first moment respectively corresponds
Business datum feature in include the historical juncture corresponding second business datum feature, the object time corresponding first
Business datum feature;Include the historical juncture corresponding the in the multiple first moment corresponding user's Figure Characteristics
The corresponding first user Figure Characteristics of two user's Figure Characteristics, the object time;The historical juncture is the object time
Last moment.
Target signature determining module 40, for being respectively corresponded according to the initially hiding feature, the multiple first moment
Business datum feature and corresponding user's Figure Characteristics of the multiple first moment, determine the target network mould
Type updates the target in duration in the feature and hides feature;The target hides feature asks for characterizing to pull with described first
Seek user's portrait information of corresponding target user.
Wherein, the target signature determining module 40 includes: history feature determination unit 401 and target signature determination unit
402;
History feature determination unit 401, for by the target network model by the second user Figure Characteristics, institute
It states the second business datum feature and the initially hiding feature is merged, and using fused initially hiding feature as institute
The history that target network model is stated in the historical juncture hides feature;
Target signature determination unit 402, for passing through the target network model for the first user Figure Characteristics, institute
It states the first business datum feature and the hiding feature of the history is merged, and fused history is hidden into feature as institute
The target that target network model is stated in the object time hides feature.
Wherein, the target network model includes the first resetting door, the second resetting door and updates door;
The target signature determination unit 402 includes: Weight Acquisition subelement 4021, and parameter obtains subelement 4022, the
One determines subelement 4023, and first determines subelement 4024, and third determines subelement 4025;
Weight Acquisition subelement 4021, for obtaining the first business datum feature in the target network model
The hiding feature of corresponding first weight parameter, corresponding second weight parameter of the first user data feature, the history is corresponding
The straggling parameter of third weight parameter and the target network model;
Parameter obtains subelement 4022, for by the first business datum feature, first weight parameter, described the
One user's Figure Characteristics, second weight parameter, the history hide feature, the third weight parameter and the deviation
Parameter exports the corresponding gate ginseng of the update door as the corresponding input parameter of activation primitive in the target network model
Several, the described corresponding gating parameter of first resetting door and the corresponding gating parameter of the second resetting door;
First determines subelement 4023, for being hidden based on the corresponding gating parameter of the first resetting door, the history
Feature and the first business datum feature determine that the first business datum feature is corresponding in the target network model
Candidate hide feature;
Second determines subelement 4024, for being hidden based on the corresponding gating parameter of the second resetting door, the history
Feature and the first user Figure Characteristics determine that the first user Figure Characteristics are corresponding in the target network model
Candidate hide feature;
Third determines subelement 4025, for according to the corresponding gating parameter of the update door, first business datum
The corresponding candidate hiding feature of feature, the corresponding candidate hiding feature of the first user Figure Characteristics and the history are hidden
Feature is hidden, determines that the target network model hides feature in the target of the object time.
Wherein, the Weight Acquisition subelement 4021, parameter obtain subelement 4022, and first determines subelement 4023, the
One determines subelement 4024, and third determines that the specific implementation of subelement 4025 can be found in embodiment corresponding to above-mentioned Fig. 6
Description to the detailed process for obtaining the hiding feature of target, will not continue to repeat here.
Wherein, the specific implementation of the history feature determination unit 401 and target signature determination unit 402 can be found in
Description in embodiment corresponding to above-mentioned Fig. 3 to the detailed process for obtaining the hiding feature of target, will not continue to go to live in the household of one's in-laws on getting married here
It states.
Wherein, the duration obtains module 10, and first obtains module 20, and second obtains module 30, and target signature determines mould
The specific implementation of block 40 can be found in the description in embodiment corresponding to above-mentioned Fig. 3 to step S101- step S104, here will
It does not continue to be repeated.
Wherein, it is to pull the corresponding first request time stamp of request by described first to draw with second that the feature, which updates duration,
It takes determined by the corresponding second request time stamp of request;Described second pulls the upper drawing that request pulls request for described first
Request is taken, described second pulls request for obtaining the multiple history recommending data information for having incidence relation with target user.
Optionally, relationship determination module 50 is triggered, for being updated in duration in the feature, the multiple history is recommended
History recommending data information in data information between the target user there are data-triggered relationship is determined as target histories
Recommending data information;
Traffic vector constructs module 60, for being based on each target histories recommending data information and the history service number
It is believed that breath, constructs the corresponding business datum feature vector of the target network model;Include in the business datum feature vector
The multiple first moment corresponding business datum feature;The corresponding target histories of one business datum feature recommend number
It is believed that breath;
Wherein, wherein the traffic vector building module 60 includes: data adding unit 601, target service determination unit
602, feature extraction unit 603, traffic vector determination unit 604;
Data adding unit 601, for each target histories recommending data information to be added to the first business trigger list,
And the first business trigger list of each target histories recommending data information will be carried as the second business trigger list;Institute
It states in the first business trigger list and the second business trigger list comprising there are history touchings between the target user
The history service data information of hair relationship;
Target service determination unit 602, in the second business trigger list, the feature to be updated in duration
Each of the first moment corresponding target histories recommending data information be determined as target service data information;
Feature extraction unit 603 is extracted for the semantic feature to the target service data information, and will be extracted
The semantic feature arrived is as the corresponding target service data characteristics of the target service data information;
Traffic vector determination unit 604, for obtain the corresponding historical data feature of the first business trigger list to
Amount, and the target service data characteristics and the historical data feature vector is spliced, and by spliced history number
It is determined as the corresponding business datum feature vector of the target network model according to feature vector.
Wherein, the data adding unit 601, target service determination unit 602, feature extraction unit 603, traffic vector
The specific implementation of determination unit 604 can be found in embodiment corresponding to above-mentioned Fig. 3 to building business datum feature vector
The description of detailed process will not continue to repeat here.
User vector constructs module 70, for passing through the corresponding target user of each target histories recommending data information
Attribute information and the corresponding historical user's attribute information of the history service data information, construct the target network model
Corresponding user's Figure Characteristics vector;It include corresponding use of the multiple first moment in user's Figure Characteristics vector
Family Figure Characteristics.
Wherein, user vector building module 70 includes: objective attribute target attribute determination unit 701, feature assembled unit 702,
User vector determination unit 703;
Objective attribute target attribute determination unit 701 was used in the second business trigger list, by each first moment
Customer attribute information corresponding to target histories recommending data information is determined as target user's attribute information;
Feature assembled unit 702, for by target user's attribute information user property feature and the target
Attention model feature in network model is combined, and the user property feature after combination is determined as the target user
The corresponding target user's Figure Characteristics of attribute information;
User vector determination unit 703, it is special for obtaining historical user associated with the first business trigger list
Vector is levied, and target user's Figure Characteristics and historical user's feature vector are spliced, and is gone through spliced
History user characteristics vector is determined as the corresponding user's Figure Characteristics vector of the target network model.
Wherein, the objective attribute target attribute determination unit 701, feature assembled unit 702, the tool of user vector determination unit 703
Body implementation can be found in the description of the detailed process in embodiment corresponding to above-mentioned Fig. 3 to building user's Figure Characteristics vector,
Here it will not continue to repeat.
Optionally, data acquisition module 80 are believed for hiding the user that feature is characterized by the target and drawing a portrait
Breath obtains multiple first candidate data information from service database;
First screening module 90, for obtaining the target user corresponding preliminary screening rule, and from the multiple the
The first candidate data information that one candidate data information sifting is consistent with the preliminary screening rule is believed as the second candidate data
Breath;
Second screening module 100, it is hidden with the target for being selected from all second candidate data information filtered out
The second candidate data information that user's Figure Characteristics vector for being characterized of hiding feature, the business datum feature vector are consistent
As data information to be recommended;
Data distributing module 110, for being issued selected data information to be recommended as target recommending data information
To the target user.
Wherein, the data distributing module 110 includes: probability prediction unit 111 and sequencing unit 112;Further, institute
State data distributing module 110 further include: predicted vector generation unit 113 clicks vector generation unit 114, optimizes subelement
115;
Probability prediction unit 111, for predicting selected multiple data to be recommended by the first clicking rate prediction model
The corresponding click probability of each of information data information to be recommended;
Sequencing unit 112, for being ranked up to the corresponding click probability of each data information to be recommended, and from row
Target recommending data information is obtained in data information to be recommended after sequence, and by the target recommending data information recommendation to described
Target user.
Optionally, predicted vector generation unit 113, for being pushed away according to every height in the multiple sub- recommending data information
The click probability for recommending data information generates the corresponding predicted vector of the first clicking rate prediction model;The predicted vector includes
Target prediction vector;The target prediction vector is for describing the subsequent time that the target user pulls request described first
The prediction probability that the first sub- recommending data information in the multiple sub- recommending data information is clicked;
Vector generation unit 114 is clicked, the target user is recorded in the subsequent time for described more for working as
When the click relationship of the second sub- recommending data information in a sub- recommending data information, described the is generated based on the click relationship
The corresponding click vector of one clicking rate prediction model;The click vector includes the target point amount of hitting to;The target point amount of hitting to
For describing the target user in the subsequent time to the second sub- recommending data in the multiple sub- recommending data information
The actual probabilities that information is clicked;
Optimize subelement 115, determines described first for the target point amount of hitting to according to the target prediction vector sum
The loss function of clicking rate prediction model, and the loss function is optimized by gradient decreasing function, and will be after optimization
Loss function be determined as objective function, and the corresponding first clicking rate prediction model of the objective function is determined as second point
Hit rate prediction model.
Wherein, the sequencing unit 112 is specifically used for based on drop ordering rule to each data information pair to be recommended
The click probability answered is ranked up, and screening clicks probability greater than click probability threshold value from the data information to be recommended after sequence
Data information to be recommended as target recommending data information, and by the target recommending data information filtered out be pushed to target use
Family.
Wherein, the probability prediction unit 111, sequencing unit 112, predicted vector generation unit 113 are clicked vector and are generated
Unit 114, the specific implementation of optimization subelement 115, which can be found in embodiment corresponding to above-mentioned Fig. 8, recommends acquisition target
The description of the detailed process of data information will not continue to repeat here.
Wherein, the triggering relationship determination module 50, traffic vector construct module 60, and user vector constructs module 70, number
According to module 80 is obtained, the specific implementation of the first screening module 90, the second screening module 100 and data distributing module 110 can
Referring to the description of the step S201- step S211 in embodiment corresponding to above-mentioned Fig. 8, will not continue to repeat here.
In embodiments of the present invention, response first pulls request for target application, obtains this and first pulls and request pair
The feature answered updates duration, this feature when updating a length of this first pull request and upper one and pull duration between request, the spy
Sign, which updates, may include multiple first moment in duration, the first moment of each of the multiple first moment is corresponding corresponding
Business datum feature and user's Figure Characteristics;So can be by target network model obtained by a upper feature updates in duration
Hiding feature be determined as initially hiding feature, so as to will initially hide feature, described more using the target network model
A first moment corresponding first user Figure Characteristics and corresponding first business number of the multiple first moment
It is merged according to feature, is exported so that the target network model can update in duration in this feature for describing target user's
The target of user's portrait information hides feature.Wherein, it can establish in continuous time period and (i.e. should by the target network model
Feature update duration and a upper feature update duration) user's Figure Characteristics and business datum feature between incidence relation, i.e.,
The target of target network model output, which hides feature, can combine the user under the different moments in continuous time period
Figure Characteristics and business datum feature are more bonded target use so as to accurately depict in business datum recommender system
User's portrait information of family hobby, and then can effectively ensure that the accuracy and completeness of user's representation data.
Further, referring to Figure 13, it is another user's representation data processing unit provided in an embodiment of the present invention
Structural schematic diagram.As shown in figure 13, which can be applied to destination server, target clothes
Business device can be the service server 2000 in embodiment corresponding to above-mentioned Fig. 1.User's representation data processing unit 1000 can
To include: processor 1001, network interface 1004 and memory 1005, in addition, user's representation data processing unit 1000
It can also include: user interface 1003 and at least one communication bus 1002.Wherein, communication bus 1002 is for realizing these
Connection communication between component.Wherein, user interface 1003 may include display screen (Display), keyboard (Keyboard), can
Selecting user interface 1003 can also include standard wireline interface and wireless interface.Network interface 1004 optionally may include mark
Wireline interface, the wireless interface (such as WI-FI interface) of standard.Memory 1004 can be high speed RAM memory, be also possible to it is non-not
Stable memory (non-volatile memory), for example, at least a magnetic disk storage.Memory 1005 optionally may be used also
To be storage device that at least one is located remotely from aforementioned processor 1001.As shown in figure 13, it stores and is situated between as a kind of computer
It may include operating system, network communication module, Subscriber Interface Module SIM and equipment control in the memory 1005 of matter using journey
Sequence.
Wherein, the network interface 1004 in user's representation data processing unit 1000 can also be with reality corresponding to above-mentioned Fig. 1
The target terminal user applied in example is connected to the network, and optional user interface 1003 can also include display screen (Display),
Keyboard (Keyboard).In user's representation data processing unit 1000 shown in Figure 13, network interface 1004 can provide network
Communication function;And user interface 1003 is mainly used for providing the interface of input for user;And processor 1001 can be used for calling
The equipment control application program stored in memory 1005, to realize:
Response pulls request for the first of target application, when acquisition described first pulls request corresponding feature update
It is long;It includes multiple first moment in duration that the feature, which updates,;
Target network model is updated obtained hiding feature in duration in a upper feature to be determined as initially hiding feature;
Obtain corresponding business datum feature of the multiple first moment from business data characteristics vector, and from
Corresponding user's Figure Characteristics of the multiple first moment are obtained in the Figure Characteristics vector of family;
According to the initially hiding feature, corresponding business datum feature of the multiple first moment and described
Multiple first moment corresponding user's Figure Characteristics determine that the target network model updates in duration in the feature
Target hides feature;The target hides feature and pulls the user for requesting corresponding target user with described first for characterizing and draw
As information.
It should be appreciated that the executable Fig. 3 above of user's representation data processing unit 1000 described in the embodiment of the present invention
Or the description in embodiment corresponding to Fig. 8 to user's representation data processing method, it also can be performed real corresponding to Figure 12 above
The description in example to user's representation data processing unit 1 is applied, details are not described herein.In addition, having to using same procedure
Beneficial effect description, is also no longer repeated.
In addition, it need to be noted that: the embodiment of the invention also provides a kind of computer storage medium, and the meter
Computer program performed by the user's representation data processing unit 1 being mentioned above is stored in calculation machine storage medium, and described
Computer program includes program instruction, when the processor executes described program instruction, is able to carry out Fig. 3 or Fig. 8 institute above
To the description of user's representation data processing method in corresponding embodiment, therefore, will no longer repeat here.In addition, right
It is described using the beneficial effect of same procedure, is also no longer repeated.It is real for computer storage medium according to the present invention
Undisclosed technical detail in example is applied, the description of embodiment of the present invention method is please referred to.
Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the program can be stored in a computer-readable storage medium
In, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, the storage medium can be magnetic
Dish, CD, read-only memory (Read-Only Memory, ROM) or random access memory (Random Access
Memory, RAM) etc..
The above disclosure is only the preferred embodiments of the present invention, cannot limit the right model of the present invention with this certainly
It encloses, therefore equivalent changes made in accordance with the claims of the present invention, is still within the scope of the present invention.
Claims (13)
1. a kind of user's representation data processing method characterized by comprising
Response pulls request for the first of target application, obtains described first and pulls the corresponding feature update duration of request;Institute
Feature is stated to update in duration comprising multiple first moment;
Target network model is updated obtained hiding feature in duration in a upper feature to be determined as initially hiding feature;
Corresponding business datum feature of the multiple first moment is obtained from business data characteristics vector, and is drawn from user
As obtaining corresponding user's Figure Characteristics of the multiple first moment in feature vector;
According to the initially hiding feature, corresponding business datum feature of the multiple first moment and the multiple
First moment corresponding user's Figure Characteristics determine that the target network model updates the target in duration in the feature
Hide feature;The target is hidden feature and is believed for characterizing to pull the user for requesting corresponding target user with described first and draw a portrait
Breath.
2. the method according to claim 1, wherein it is to pull request by described first that the feature, which updates duration,
Corresponding first request time stamp pulls determined by corresponding second request time of request stabs with second;Described second pulls and asks
It asks and pulls upper the one of request for described first and pull request, described second pulls request for obtaining with target user with being associated with
Multiple history recommending data information of relationship;
The method also includes:
It is updated in duration in the feature, by there are numbers between the target user in the multiple history recommending data information
History recommending data information according to triggering relationship is determined as target histories recommending data information;
Based on each target histories recommending data information and history service data information, it is corresponding to construct the target network model
Business datum feature vector;It include corresponding business number of the multiple first moment in the business datum feature vector
According to feature;The corresponding target histories recommending data information of one business datum feature;
Pass through each corresponding target user's attribute information of target histories recommending data information and the history service number
It is believed that ceasing corresponding historical user's attribute information, the corresponding user's Figure Characteristics vector of the target network model is constructed;It is described
It include corresponding user's Figure Characteristics of the multiple first moment in user's Figure Characteristics vector.
3. according to the method described in claim 2, it is characterized in that, it is described based on each target histories recommending data information and
History service data information constructs the corresponding business datum feature vector of the target network model, comprising:
Each target histories recommending data information is added to the first business trigger list, and each target histories will be carried
First business trigger list of recommending data information is as the second business trigger list;The first business trigger list and described
Comprising there are the history service data informations of history triggering relationship between the target user in second business trigger list;
In the second business trigger list, by the first moment corresponding target histories each of in feature update duration
Recommending data information is determined as target service data information;
The semantic feature of the target service data information is extracted, and using the semantic feature extracted as the target
The corresponding target service data characteristics of service data information;
Obtain the corresponding historical data feature vector of the first business trigger list, and by the target service data characteristics with
The historical data feature vector is spliced, and spliced historical data feature vector is determined as the target network mould
The corresponding business datum feature vector of type.
4. according to the method described in claim 3, it is characterized in that, described pass through each target histories recommending data information
Corresponding target user's attribute information and the corresponding historical user's attribute information of the history service data information construct institute
State the corresponding user's Figure Characteristics vector of target network model, comprising:
It, will be corresponding to the target histories recommending data information at each first moment in the second business trigger list
Customer attribute information is determined as target user's attribute information;
User property feature in target user's attribute information and the attention model in the target network model is special
Sign is combined, and the user property feature after combination is determined as the corresponding target user of target user's attribute information and is drawn
As feature;
Historical user's feature vector associated with the first business trigger list is obtained, and target user portrait is special
Sign is spliced with historical user's feature vector, and spliced historical user's feature vector is determined as the target network
The corresponding user's Figure Characteristics vector of network model.
5. the method according to claim 1, wherein when the multiple first moment includes historical juncture and target
It carves, includes the historical juncture corresponding second business datum in the multiple first moment corresponding business datum feature
The corresponding first business datum feature of feature, the object time;The multiple first moment corresponding user draws a portrait special
It is special comprising the historical juncture corresponding second user Figure Characteristics, the corresponding first user portrait of the object time in sign
Sign;The historical juncture is the last moment of the object time;
It is described according to the initially hiding feature, corresponding business datum feature of the multiple first moment and described
Multiple first moment corresponding user's Figure Characteristics determine that the target network model updates in duration in the feature
Target hides feature, comprising:
By the target network model by the second user Figure Characteristics, the second business datum feature and it is described just
The hiding feature that begins is merged, and using fused initially hiding feature as the target network model in the historical juncture
History hide feature;
By the target network model by the first user Figure Characteristics, the first business datum feature and described go through
History is hidden feature and is merged, and fused history is hidden feature as the target network model in the object time
Target hide feature.
6. according to the method described in claim 5, it is characterized in that, the target network model includes the first resetting door, second
It resets door and updates door;
It is described by the target network model by the first user Figure Characteristics, the first business datum feature and institute
It states the hiding feature of history to be merged, and fused history is hidden into feature as the target network model in the target
The target at moment hides feature, comprising:
In the target network model, corresponding first weight parameter of the first business datum feature, the first user are obtained
Corresponding second weight parameter of data characteristics, the history hide the corresponding third weight parameter of feature and the target network
The straggling parameter of network model;
By the first business datum feature, first weight parameter, the first user Figure Characteristics, second weight
Parameter, the history hide feature, the third weight parameter and the straggling parameter as in the target network model
The corresponding input parameter of activation primitive, export the corresponding gating parameter of the updates door, described first reset the corresponding door of door
Control parameter and the corresponding gating parameter of the second resetting door;
Based on the corresponding gating parameter of the first resetting door, the history hides feature and first business datum is special
Sign determines the corresponding candidate hiding feature of the first business datum feature in the target network model;
Feature is hidden based on the corresponding gating parameter of the second resetting door, the history and first user portrait is special
Sign determines the corresponding candidate hiding feature of the first user Figure Characteristics in the target network model;
According to the corresponding candidate hiding feature of the corresponding gating parameter of the update door, the first business datum feature, described
The corresponding candidate hiding feature of first user's Figure Characteristics and the history hide feature, determine the target network model
Feature is hidden in the target of the object time.
7. the method according to claim 1, wherein further include:
The user portrait information that feature is characterized is hidden by the target, and multiple first times are obtained from service database
Select data information;
Obtain the target user corresponding preliminary screening rule, and from the multiple first candidate data information sifting with it is described
The first candidate data information that preliminary screening rule is consistent is as the second candidate data information;
It selects to hide the user picture that feature is characterized with the target from all second candidate data information filtered out
The second candidate data information being consistent as feature vector, the business datum feature vector is as data information to be recommended;
The target user is handed down to using selected data information to be recommended as target recommending data information.
8. the method according to the description of claim 7 is characterized in that described using selected data information to be recommended as target
Recommending data information is handed down to the target user, comprising:
Each of selected data information to be recommended data information pair to be recommended is predicted by the first clicking rate prediction model
The click probability answered;
The corresponding click probability of each data information to be recommended is ranked up, and from the data information to be recommended after sequence
Middle acquisition target recommending data information, and give the target recommending data information recommendation to the target user.
9. according to the method described in claim 8, it is characterized in that, described to the corresponding point of each data information to be recommended
It hits probability to be ranked up, and obtains target recommending data information from the data information to be recommended after sequence, and by the target
Recommending data information recommendation gives the target user, comprising:
Based on drop ordering rule the click probability of each data information to be recommended is ranked up, and from after sequence wait push away
It recommends in data information screening and clicks probability and be greater than and click the data information to be recommended of probability threshold value as target recommending data information,
And the target recommending data information filtered out is pushed to target user.
10. according to the method described in claim 8, it is characterized in that, the target recommending data information includes multiple sub- recommendations
Data information;
The method also includes:
Generate at described first point according to the click probability of every sub- recommending data information in the multiple sub- recommending data information
Hit the corresponding predicted vector of rate prediction model;The predicted vector includes target prediction vector;The target prediction vector is used for
It describes the target user and pulls the subsequent time of request to first in the multiple sub- recommending data information described first
The prediction probability that sub- recommending data information is clicked;
It is pushed away in the subsequent time for the second son in the multiple sub- recommending data information when the target user is recorded
When recommending the click relationship of data information, the corresponding point of the first clicking rate prediction model is generated based on the click relationship and is hit to
Amount;The click vector includes the target point amount of hitting to;The target point amount of hitting to is for describing the target user under described
The actual probabilities that one moment clicked the second sub- recommending data information in the multiple sub- recommending data information;
The loss letter of the first clicking rate prediction model is determined according to the target point amount of hitting to described in the target prediction vector sum
Number, and the loss function is optimized by gradient decreasing function, and the loss function after optimization is determined as target letter
Number, and the corresponding first clicking rate prediction model of the objective function is determined as the second clicking rate prediction model.
11. a kind of user's representation data processing unit characterized by comprising
Duration obtains module, pulls request for the first of target application for responding, and obtains described first and pulls request correspondence
Feature update duration;It includes multiple first moment in duration that the feature, which updates,;
First obtain module, for using target network model a upper feature update duration in obtained hiding feature as at the beginning of
Begin to hide feature;
Second obtains module, for obtaining corresponding business number of the multiple first moment from business data characteristics vector
According to feature, and corresponding user's Figure Characteristics of the multiple first moment are obtained from user's Figure Characteristics vector;
Target signature determining module, for according to the initially hiding feature, corresponding business of the multiple first moment
Data characteristics and corresponding user's Figure Characteristics of the multiple first moment, determine the target network model in institute
It states the target that feature updates in duration and hides feature;The target hides feature, and for characterizing, with described first to pull request corresponding
Target user user draw a portrait information.
12. a kind of user's representation data processing unit, described device is applied to computer equipment characterized by comprising processing
Device, memory and network interface;
The processor is connected with memory, network interface, wherein network interface is described to deposit for providing data communication function
Reservoir is for storing program code, and the processor is for calling said program code, to execute as claim 1-10 is any
Method described in.
13. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has computer journey
Sequence, the computer program include program instruction, and described program instructs when being executed by a processor, execute such as claim 1-10
Described in any item methods.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910312225.3A CN110162698B (en) | 2019-04-18 | 2019-04-18 | User portrait data processing method, device and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910312225.3A CN110162698B (en) | 2019-04-18 | 2019-04-18 | User portrait data processing method, device and storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110162698A true CN110162698A (en) | 2019-08-23 |
CN110162698B CN110162698B (en) | 2023-10-20 |
Family
ID=67639463
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910312225.3A Active CN110162698B (en) | 2019-04-18 | 2019-04-18 | User portrait data processing method, device and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110162698B (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110598016A (en) * | 2019-09-11 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Method, device, equipment and medium for recommending multimedia information |
CN110909040A (en) * | 2019-11-08 | 2020-03-24 | 支付宝(杭州)信息技术有限公司 | Business delivery auxiliary method and device and electronic equipment |
CN110929206A (en) * | 2019-11-20 | 2020-03-27 | 腾讯科技(深圳)有限公司 | Click rate estimation method and device, computer readable storage medium and equipment |
CN111400603A (en) * | 2020-03-20 | 2020-07-10 | 腾讯科技(深圳)有限公司 | Information pushing method, device and equipment and computer readable storage medium |
CN111538751A (en) * | 2020-03-23 | 2020-08-14 | 重庆特斯联智慧科技股份有限公司 | Tagged user portrait generation system and method for Internet of things data |
CN111782959A (en) * | 2020-07-20 | 2020-10-16 | 深圳市欢太科技有限公司 | User portrait updating method and device, computer equipment and storage medium |
CN112132622A (en) * | 2020-09-25 | 2020-12-25 | 北京达佳互联信息技术有限公司 | Data estimation method and device |
CN112541123A (en) * | 2020-12-24 | 2021-03-23 | 北京百度网讯科技有限公司 | Map region recommendation method, device, equipment and storage medium |
WO2021159393A1 (en) * | 2020-02-13 | 2021-08-19 | 深圳市欢太科技有限公司 | Information pushing method and apparatus, and server and storage medium |
CN113515696A (en) * | 2021-05-17 | 2021-10-19 | 上海众源网络有限公司 | Recommendation method and device, electronic equipment and storage medium |
CN114048392A (en) * | 2022-01-13 | 2022-02-15 | 北京达佳互联信息技术有限公司 | Multimedia resource pushing method and device, electronic equipment and storage medium |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105205184A (en) * | 2015-10-30 | 2015-12-30 | 云南大学 | Latent variable model-based user preference extraction method |
WO2017181612A1 (en) * | 2016-04-18 | 2017-10-26 | 乐视控股(北京)有限公司 | Personalized video recommendation method and device |
CN107423442A (en) * | 2017-08-07 | 2017-12-01 | 火烈鸟网络(广州)股份有限公司 | Method and system, storage medium and computer equipment are recommended in application based on user's portrait behavioural analysis |
CN107644047A (en) * | 2016-07-22 | 2018-01-30 | 华为技术有限公司 | Tag Estimation generation method and device |
US20180218431A1 (en) * | 2017-01-31 | 2018-08-02 | Wal-Mart Stores, Inc. | Providing recommendations based on user-generated post-purchase content and navigation patterns |
CN108898429A (en) * | 2018-06-19 | 2018-11-27 | 平安科技(深圳)有限公司 | Electronic device, preference tendency prediction technique and computer readable storage medium |
CN109145210A (en) * | 2018-08-17 | 2019-01-04 | 腾讯科技(深圳)有限公司 | A kind of method, apparatus and storage medium of information recommendation |
-
2019
- 2019-04-18 CN CN201910312225.3A patent/CN110162698B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105205184A (en) * | 2015-10-30 | 2015-12-30 | 云南大学 | Latent variable model-based user preference extraction method |
WO2017181612A1 (en) * | 2016-04-18 | 2017-10-26 | 乐视控股(北京)有限公司 | Personalized video recommendation method and device |
CN107644047A (en) * | 2016-07-22 | 2018-01-30 | 华为技术有限公司 | Tag Estimation generation method and device |
US20180218431A1 (en) * | 2017-01-31 | 2018-08-02 | Wal-Mart Stores, Inc. | Providing recommendations based on user-generated post-purchase content and navigation patterns |
CN107423442A (en) * | 2017-08-07 | 2017-12-01 | 火烈鸟网络(广州)股份有限公司 | Method and system, storage medium and computer equipment are recommended in application based on user's portrait behavioural analysis |
CN108898429A (en) * | 2018-06-19 | 2018-11-27 | 平安科技(深圳)有限公司 | Electronic device, preference tendency prediction technique and computer readable storage medium |
CN109145210A (en) * | 2018-08-17 | 2019-01-04 | 腾讯科技(深圳)有限公司 | A kind of method, apparatus and storage medium of information recommendation |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110598016A (en) * | 2019-09-11 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Method, device, equipment and medium for recommending multimedia information |
CN110909040A (en) * | 2019-11-08 | 2020-03-24 | 支付宝(杭州)信息技术有限公司 | Business delivery auxiliary method and device and electronic equipment |
CN110909040B (en) * | 2019-11-08 | 2022-03-04 | 支付宝(杭州)信息技术有限公司 | Business delivery auxiliary method and device and electronic equipment |
CN110929206A (en) * | 2019-11-20 | 2020-03-27 | 腾讯科技(深圳)有限公司 | Click rate estimation method and device, computer readable storage medium and equipment |
CN110929206B (en) * | 2019-11-20 | 2023-04-07 | 腾讯科技(深圳)有限公司 | Click rate estimation method and device, computer readable storage medium and equipment |
WO2021159393A1 (en) * | 2020-02-13 | 2021-08-19 | 深圳市欢太科技有限公司 | Information pushing method and apparatus, and server and storage medium |
CN111400603A (en) * | 2020-03-20 | 2020-07-10 | 腾讯科技(深圳)有限公司 | Information pushing method, device and equipment and computer readable storage medium |
CN111538751B (en) * | 2020-03-23 | 2021-05-04 | 重庆特斯联智慧科技股份有限公司 | Tagged user portrait generation system and method for Internet of things data |
CN111538751A (en) * | 2020-03-23 | 2020-08-14 | 重庆特斯联智慧科技股份有限公司 | Tagged user portrait generation system and method for Internet of things data |
CN111782959A (en) * | 2020-07-20 | 2020-10-16 | 深圳市欢太科技有限公司 | User portrait updating method and device, computer equipment and storage medium |
CN111782959B (en) * | 2020-07-20 | 2023-07-14 | 深圳市与飞科技有限公司 | User portrait updating method, device, computer equipment and storage medium |
CN112132622A (en) * | 2020-09-25 | 2020-12-25 | 北京达佳互联信息技术有限公司 | Data estimation method and device |
CN112541123A (en) * | 2020-12-24 | 2021-03-23 | 北京百度网讯科技有限公司 | Map region recommendation method, device, equipment and storage medium |
CN113515696A (en) * | 2021-05-17 | 2021-10-19 | 上海众源网络有限公司 | Recommendation method and device, electronic equipment and storage medium |
CN114048392A (en) * | 2022-01-13 | 2022-02-15 | 北京达佳互联信息技术有限公司 | Multimedia resource pushing method and device, electronic equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN110162698B (en) | 2023-10-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110162698A (en) | A kind of user's representation data processing method, device and storage medium | |
CN105959374B (en) | A kind of data recommendation method and its equipment | |
CN109597844B (en) | Core user mining method and system based on deep neural network and graph network | |
CN108108821A (en) | Model training method and device | |
CN110008409A (en) | Based on the sequence of recommendation method, device and equipment from attention mechanism | |
CN109862432A (en) | Clicking rate prediction technique and device | |
CN109902708A (en) | A kind of recommended models training method and relevant apparatus | |
CN110309427A (en) | A kind of object recommendation method, apparatus and storage medium | |
WO2021027256A1 (en) | Method and apparatus for processing interactive sequence data | |
CN107423442A (en) | Method and system, storage medium and computer equipment are recommended in application based on user's portrait behavioural analysis | |
CN110032630A (en) | Talk about art recommendation apparatus, method and model training equipment | |
CN108351986A (en) | Learning system, learning device, learning method, learning program, training data generating means, training data generation method, training data generate program, terminal installation and threshold value change device | |
CN110263244A (en) | Content recommendation method, device, storage medium and computer equipment | |
CN110378434A (en) | Training method, recommended method, device and the electronic equipment of clicking rate prediction model | |
CN106951435A (en) | News recommends method, equipment and programmable device | |
CN109960761A (en) | Information recommendation method, device, equipment and computer readable storage medium | |
CN109816438B (en) | Information pushing method and device | |
CN111400603A (en) | Information pushing method, device and equipment and computer readable storage medium | |
CN111241394A (en) | Data processing method and device, computer readable storage medium and electronic equipment | |
CN109829064A (en) | Media resource shares and playing method and device, storage medium and electronic device | |
CN109766925A (en) | Feature fusion, device, electronic equipment and storage medium | |
CN110110372A (en) | A kind of user's timing behavior automatic segmentation prediction technique | |
CN110019650A (en) | Method, apparatus, storage medium and the electronic equipment of search associational word are provided | |
CN110263136B (en) | Method and device for pushing object to user based on reinforcement learning model | |
CN109978575A (en) | A kind of method and device excavated customer flow and manage scene |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |