CN110162700A - The training method of information recommendation and model, device, equipment and storage medium - Google Patents
The training method of information recommendation and model, device, equipment and storage medium Download PDFInfo
- Publication number
- CN110162700A CN110162700A CN201910326702.1A CN201910326702A CN110162700A CN 110162700 A CN110162700 A CN 110162700A CN 201910326702 A CN201910326702 A CN 201910326702A CN 110162700 A CN110162700 A CN 110162700A
- Authority
- CN
- China
- Prior art keywords
- article
- user
- vector
- sample
- different types
- 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.)
- Pending
Links
Abstract
The embodiment of the present invention provides a kind of information recommendation method and device, the training method of information recommendation model and device and storage medium, and information recommendation method includes: the recommendation request for receiving the correspondence target user of client;The corresponding different types of user characteristics of the target user are extracted, the user vector to form the target user is combined according to the different types of user characteristics;The different types of article characteristics for extracting article, the article vector to form the article is combined according to the different types of article characteristics;According to the distance between article vector and user vector, the determining vector distance with the user vector meets the article of condition;Corresponding recommendation information is sent to the client based on the article for the condition that meets.The embodiment of the present invention can all-sidedly and accurately carry out personalized recommendation.
Description
Technical field
The present invention relates to technical field of information processing, in particular to a kind of information recommendation method and device, information recommendation mould
Training method and device, the computer equipment and storage medium of type.
Background technique
It is increasingly developed with information technology, information recommendation have become one of current application of net it is important in
Hold.For example, browsing news in user constantly recommends possible interested content, recommend tendency purchase when user browses commodity
Commodity, etc..
The customized information of study user is the core of information recommendation model, just can be carried out the accurate recommendation of information in this way.
Although, all there is limitation in terms of personalized recommendation in the information recommendation types of models multiplicity that the relevant technologies provide.
Summary of the invention
The embodiment of the present invention provide a kind of information recommendation method and device, the training method of information recommendation model and device,
Computer equipment and storage medium can all-sidedly and accurately carry out personalized recommendation.
The technical solution of the embodiment of the present invention is achieved in that
The embodiment of the present invention provides a kind of information recommendation method, comprising: receives the recommendation of the correspondence target user of client
Request;The corresponding different types of user characteristics of the target user are extracted, are combined according to the different types of user characteristics
Form the user vector of the target user;The different types of article characteristics for extracting article, according to the different types of object
Product feature combines the article vector to form the article;According to the distance between the article vector and the user vector, really
The fixed vector distance with the user vector meets the article of condition;The article based on the condition that meets is sent out to the client
Send corresponding recommendation information.
The embodiment of the present invention provides a kind of training method of information recommendation model, comprising: obtains training sample set, the instruction
The different type feature and the sample of users that training sample in white silk sample set includes sample of users are for the article
Practical scoring;By the first combination layer of the information recommendation model, by the different types of user characteristics of the sample of users
Coding vector be combined, obtain the user vector of sample of users;It, will by the second combination layer of the information recommendation model
The coding vector of the different types of article characteristics of the article is combined, and obtains the article vector of sample article;Pass through institute
That states information recommendation model determines the sample of users according to the vector distance between the user vector and the sample article
Prediction scoring to the sample article;According to prediction scoring and information recommendation described in the error update actually to score
The parameter of model, until the loss function of the information recommendation model is restrained.
The embodiment of the present invention provides a kind of information recommending apparatus, comprising: receiving module, for receiving the correspondence mesh of client
Mark the recommendation request of user;User vector module, for extracting the corresponding different types of user characteristics of the target user, root
The user vector to form the target user is combined according to the different types of user characteristics;Article vector module, for extracting
The different types of article characteristics of article, combined according to the different types of article characteristics to be formed the article of the article to
Amount;Enquiry module is used for according to the distance between the article vector and the user vector, determining and the user vector
Vector distance meets the article of condition;Recommending module, for the article based on the condition that meets to the client send pair
The recommendation information answered.
The embodiment of the present invention provides a kind of training device of information recommendation model, comprising: sample module, for obtaining training
Sample set, the training sample that the training sample is concentrated include the different type feature and the sample of users of sample of users
For the practical scoring of the article;Subscriber-coded module, for passing through the first combination layer of the information recommendation model, by institute
The coding vector for stating the different types of user characteristics of sample of users is combined, and obtains the user vector of sample of users;Article
Coding module, for passing through the second combination layer of the information recommendation model, by the different types of article characteristics of the article
Coding vector be combined, obtain the article vector of sample article;Grading module, for passing through the information recommendation model
Determine the sample of users to the sample article according to the vector distance between the user vector and the sample article
Prediction scoring;Module is restrained, for according to prediction scoring and information recommendation mould described in the error update actually to score
The parameter of type, until the loss function of the information recommendation model is restrained.
The embodiment of the present invention provides a kind of computer equipment, including processor and can run on a processor for storing
Computer program memory;Wherein, the processor is for when running the computer program, realizing any reality of the present invention
It applies information recommendation method provided by example or realizes the training side of information recommendation model provided by any embodiment of the present invention
Method.
The embodiment of the present invention provides a kind of storage medium, is stored thereon with computer program, and the computer program is processed
Information recommendation method provided by any embodiment of the present invention is realized when device executes or realizes that any embodiment of the present invention is mentioned
The training method of the information recommendation model of confession.
For the different types of user characteristics of target user in the embodiment of the present invention, it is corresponding that user is formed by combination
User vector forms the corresponding article vector of article by combination, due to respectively for the different types of article characteristics of article
Learn the comprehensive feature to target user and article, thus allow for fully personalized recommendation, is based on different type
Feature combines to form corresponding vector, and meets the article of condition according to the mode of vector distance, and recommendation can be effectively reduced
Computation complexity in the process, to realize good recommendation efficiency.
Detailed description of the invention
Fig. 1 is the system architecture diagram of information recommendation method in the embodiment of the present invention;
Fig. 2 is the process timing diagram of information recommendation method in the embodiment of the present invention;
Fig. 3 is the architecture diagram of information recommendation model in the embodiment of the present invention;
Fig. 4 is one optional application scenarios schematic diagram of information recommendation method in the embodiment of the present invention;
Fig. 5 is that terminal receives the interface schematic diagram after recommendation information in the embodiment of the present invention;
Fig. 6 is the structural schematic diagram of information recommending apparatus in the embodiment of the present invention;
Fig. 7 is the structural schematic diagram of the training device of information recommendation model in the embodiment of the present invention;
Fig. 8 is the flow diagram of information recommendation method in the embodiment of the present invention;
Fig. 9 is the schematic diagram of full Connection Neural Network in the embodiment of the present invention;
Figure 10 is the schematic diagram encoded in the embodiment of the present invention based on Code Mapping Tables;
Figure 11 is the flow diagram of the training method of information recommendation model in the embodiment of the present invention;
Figure 12 is the original being adjusted using attention mechanism to the weight of different types of feature in the embodiment of the present invention
Manage schematic diagram;
Figure 13 is the flow diagram of the training method of information recommendation model in the optional illustrative examples of the present invention;
Figure 14 is the architecture diagram of information recommendation model in the optional illustrative examples of the present invention;
Figure 15 is the training frame diagram of information recommendation model in the embodiment of the present invention;
Figure 16 is the schematic diagram that feature is carried out vectorization by information recommendation model in the embodiment of the present invention;
Figure 17 is the flow diagram of information recommendation method in the optional illustrative examples of the present invention.
Specific embodiment
The present invention is further described in detail below with reference to the accompanying drawings and embodiments.It should be appreciated that described herein
Specific embodiment is only used to explain the present invention, is not intended to limit the present invention.
Unless otherwise defined, all technical and scientific terms used herein and belong to technical field of the invention
The normally understood meaning of technical staff is identical.Term as used herein in the specification of the present invention is intended merely to description tool
The purpose of the embodiment of body, it is not intended that in the limitation present invention.Term " and or " used herein includes one or more phases
Any and all combinations of the listed item of pass.
Before the present invention will be described in further detail, noun involved in the embodiment of the present invention and term are said
Bright, noun involved in the embodiment of the present invention and term are suitable for following explanation.
1) article herein refers to for various articles (virtual objects and physical item) recommended to the user and services (example
Such as social interaction server, article, video media services etc.).
2) user characteristics, the feature analyzed according to the log recording of user or attribute, the user characteristics root of each type
Different groupings can be divided into again according to physical meaning.E.g., including following kind of user characteristics: user base attribute information
(including following grouping: age, gender etc.), user draw a portrait information (including following grouping: the classification of user, user label
Deng), (including the grouping below: acts and efforts for expediency characteristic information, long-term action characteristic information, such as short of user behavior characteristics information
Phase/long-term classification, the label etc. for reading article).
3) training sample includes the text of user characteristics and article characteristics, such as information recommendation system is pushed away to user
The recommendation log recording of article is recommended, the collection of training sample is collectively referred to as training dataset.
4) loss function (Loss Function) is also cost function (cost function), is Neural Network Optimization
Objective function.
5) neural network (Neural Networks, NN) is by a large amount of, simple processing unit (referred to as neuron)
The complex networks system for widely interconnecting and being formed, it reflects many essential characteristics of human brain function, is a height
Complicated non-linear dynamic learning system.
6) it recalls, is concentrated in the article to be recommended of information recommendation system, a batch selected according to user characteristics is wait push away
The article recommended.
7) it sorts, priority is determined according to the condition of setting for the article recalled in information recommendation system, wherein preferentially
High (the sorting preceding) article of grade will be by preferential recommendation to user.
In the way of recommendation being currently known, essence is to rely on the behavioural analysis to the historical viewings article of user itself
Determining, the interest extension of user cannot be supported well, cannot accomplish personalized accurate recommendation.
In order to realize more personalized accurate recommendation, the embodiment of the present invention provides information recommendation method and information recommendation mould
The training method of type, the information recommending apparatus of implementation information recommended method, implementation information recommended models training method training
Device, computer equipment and the executable journey for storing the training method for realizing information recommendation method or information recommendation model
The storage medium of sequence.
For the implementation of information recommendation method, the embodiment of the present invention provides the side of terminal side or server side implementation
Case will be illustrated the exemplary implement scene of information recommendation.As shown in Figure 1, the information provided for one embodiment of the invention
The architecture diagram of the system of recommended method, including terminal 100, server 200 and the database 300 connected based on network communication.With
Family by the client for the Information application installed in terminal 100, open, log in or the client of click information application after can be with
Information page is browsed, realizes that the information recommending apparatus of information recommendation method provided in an embodiment of the present invention can be used as Information application
Serve end program be installed on server 200.Please refer to Fig. 2, information recommendation method includes: S1, and server 200 is from data
It is obtained in library 300 and recommends log recording;S2,200 pairs of recommendation log recordings of server carry out feature extraction, coding, cleaning, with shape
At training sample set;S3, server 200 pass through training sample set training information recommended models;S4, server 200 establish full dose
The index data base of article;Wherein, full dose object is determined according to recommendation log recording by the information recommendation model after training
The corresponding article vector of product establishes the index data of full dose article according to the corresponding relationship of the identity of article and article vector
Library.S5, terminal 100 send recommendation request to server 200;Wherein, when the terminal 100 for the client for being equipped with Information application is examined
User is measured to open, log in or when click information applications client, then send recommendation request.S6, server 200 determine the mesh
Mark the corresponding user vector of user;Wherein, server 200 passes through the corresponding mesh of information recommendation model extraction client after training
The different types of user characteristics for marking user, determine the corresponding user vector of the target user;S7 determines corresponding article,
Corresponding recommendation information is sent to the client according to the article;Wherein, server 200 is according to user vector and full dose object
The distance between article vector in the index data base of product, determines corresponding article, and according to the article to the client
Send corresponding recommendation information;S8 receives recommendation information and chooses;Wherein, user can pass through the client of Information application
The specify information page views the recommendation information that information recommendation system is sent to client, and is chosen according to itself hobby.S9, shape
It stores at recommendation log recording into database;Wherein, server 200 recommends recommendation information and the user of article to user
The operation for choosing the article recommended, which is correspondingly formed, recommends log recording to store into database.
Wherein, the different types of user characteristics for extracting the corresponding target user of client, determine the target user couple
The information recommendation model for the user vector answered can be obtained by being formed after training sample is trained with history recommendation record.
Referring to Fig. 3, the information recommendation model is successively for the configuration diagram of information recommendation model provided by one embodiment of the invention
Including coding layer 50, splicing layer 51, dimensionality reduction layer 52 and prediction interval 53.The coding layer 50, for different types of from extracting
Feature, and extracted feature is encoded, corresponding coding vector is obtained, optionally, which can be
Embedding layers.The splicing layer 51, for the coding vector of different types of feature to be carried out splicing, the dimensionality reduction
Layer 52 obtains corresponding user vector for carrying out dimension-reduction treatment to spliced coding vector.Wherein, coding layer 50, splicing
Layer 51, dimensionality reduction layer 52 respectively include being formed two groups of independent network portion, during being trained to information recommendation model,
Can form training sample according to history recommendation record, two independent network portions for respectively in training sample to
Family feature and article characteristics carry out above-mentioned processing.The prediction interval 53, for receiving the coding vector of two dimensionality reduction layers 52 output,
Determine that user scores to the prediction of article according to the vector distance between two vectors, according to prediction scoring and corresponding trained sample
The error of practical scoring in this updates the parameter of information recommendation model.Optionally, prediction interval 53 can score according to prediction
With sample of users in corresponding training sample for article practical scoring as error signal, in information recommendation model by instead
To propagation, in each network layer of backpropagation, in the way of the solution of various gradients, determine loss function relative to network
The parameter of the network layer is subtracted corresponding gradient and realizes update by the gradient of layer parameter.During information recommendation, Ke Yitong
The different types of user characteristics of information recommendation model extraction target user after crossing the training, and by corresponding to user characteristics
The independent network portion handled is successively encoded, is spliced and dimensionality reduction, obtain user corresponding with target user to
Amount.
Wherein, the Information application can be the application program that information is browsed for user under any implement scene, such as
All kinds of social applications, including microblogging, wechat, trill etc.;Search engine application, including mobile phone Baidu etc.;News application, including one
Point information, today's tops etc.;Online store application, including Meituan, Taobao etc..Information page refers in Information application for user
Browse the software interface of information.Server can be independent information page according to corresponding article information recommended to the user and show
Show, the partial page region for being also possible in any information page of user's browsing include is shown.
For it can recommend the Information application A of article and video to user, here, article refers to text recommended to the user
Chapter and/or video.The client of user mount message application A in the terminal.Server by obtain Information application A to user into
The recommendation log recording that row article is recommended, by being based on the article from the article characteristics for extracting article in log recording are recommended
Feature determines the corresponding article vector of article, and the index data of full dose article is established according to the corresponding relationship of article and article vector
Library, referring to Fig. 4, an optional application scenarios schematic diagram of the information recommendation method provided for one embodiment of the invention,
In, the index data base of full dose article includes article index database and video index database.Information application A carries out information and pushes away
The implementation process recommended, which specifically includes that, establishes article index library and/or video index database;Recalled by article/video recalls
Module recalls relative article to user;Then by article, slightly row/video is slightly arranged again;Article essence row/video essence row's output article
And/or the ranking results of video, finally the ranking results of article and video are sorted again by mixing model, it finally will be by mixed
The personalization results that row's model sorts again recommend user.Here, the article to be recommended referred in information recommendation system is recalled
It concentrates, a batch selected according to user characteristics article to be recommended.Thick row, which refers to, recalls the degree of correlation according to user's Long-term Interest
Higher article, while mitigating essence row's staged pressure.Essence row refers to that basis slightly arranges the list of articles recalled, good by off-line training
Order models predict clicking rate (CTR), finally issue the preceding multiple articles of sequence as essence row and export result.Mixing refers to
Different ratios is set for essence row's output result of different articles, hybrid-sorting determination is carried out according to the article of different proportion and is pushed away
Recommend result.
Information recommendation method provided by the embodiment of the present invention can be applied in server 200 as shown in Figure 1, be used in combination
In realize the article in process shown in Fig. 4 recall/video recalls the node that module recalls to user relative article.When user beats
When being used to receive the specify information page for recommending article of Information application A is opened, terminal 100 sends recommendation request to server 200,
Server 200 extracts the user characteristics of the user according to recommendation request, can be from recommendation log recording etc. for old user
Corresponding user characteristics are got in behavior record;For new user, it is available to user installation, register, log in this
The behavior record of Information application A and the user characteristics for extracting user, wherein extracted user characteristics may include but unlimited
In with Types Below: primary attribute information, acts and efforts for expediency feature, interest attribute feature etc., according to the physics of the feature of each type
Meaning is grouped, and e.g., user base attribute information may include following grouping: age, gender;User's portrait information can wrap
Include following grouping: the classification of user, user label;User behavior characteristics information may include grouping below: acts and efforts for expediency
Characteristic information, long-term action characteristic information, short-term/long-term classification, the label etc. for reading article.Server passes through according to inhomogeneity
The user characteristics of type carry out merging the corresponding user vector of determining user, according between user vector and each article vector away from
From, determine article corresponding to the user, and screened according to the screening conditions of setting based on the article, it is as thick in successively passed through
After row, essence row or mixing, corresponding recommendation information is pushed by the specify information page user oriented of the Information application A.It please join
Fig. 5 is read, after the recommendation information returned for server 200 to terminal 100, the interface schematic diagram of terminal 100, target user is by beating
Open terminal 100 for receiving recommendation information interface, can be " selected " interface in " having a look at " of wechat application, such as to obtain
Get the corresponding recommendation information of the article for meeting condition with the user vector of the target user of the transmission of server 200.
Information recommendation method provided in an embodiment of the present invention can be implemented using server side, just implement the information recommendation side
For the hardware configuration of the information recommending apparatus of method, referring to Fig. 6, being the one of information recommending apparatus provided in an embodiment of the present invention
A optional hardware structural diagram, the information recommending apparatus include: at least one first processor 301, first memory
302, at least one first network interface 304 and the first user interface 306.Various components in information recommending apparatus pass through bus
System is coupled.It should be understood that bus system is for realizing the connection communication between these components.Bus system is except packet
It includes except data/address bus, further includes power bus, control bus and status signal bus in addition.But for the sake of clear explanation,
Various buses are all designated as bus system in Fig. 3.
Wherein, the first user interface 306 may include display, keyboard, mouse, trace ball, click wheel, key, button,
Touch-sensitive plate or touch screen etc..
It is appreciated that first memory 302 can be volatile memory or nonvolatile memory, may also comprise volatile
Both property and nonvolatile memory.The memory of description of the embodiment of the present invention is intended to include the memory of any suitable type.
First memory 302 in the embodiment of the present invention is for storing various types of data to support information recommending apparatus
Operation.The example of these data includes: any executable program for operating on information recommending apparatus, such as operating system
And application program;Training sample;Recommend log recording;The index data base etc. of full dose article;Wherein, operating system includes various
System program, for example, ccf layer, core library layer, driving layer etc., hardware based for realizing various basic businesses and processing
Task.Application program may include various application programs, such as media player (Media Player), browser
(Browser) etc., for realizing various applied business.Realize the information recommendation of information recommendation method provided in an embodiment of the present invention
Device may include in the application.
The method that the embodiments of the present invention disclose can be applied in first processor 301, or by first processor
301 realize.First processor 301 may be a kind of IC chip, the processing capacity with signal.During realization,
Each step of the above method can pass through the integrated logic circuit of the hardware in first processor 301 or the instruction of software form
It completes.Above-mentioned first processor 301 can be general processor, digital signal processor (DSP, Digital Signal
Processor) either other programmable logic device, discrete gate or transistor logic, discrete hardware components etc..The
One processor 301 may be implemented or execute disclosed each method, step and logic diagram in the embodiment of the present invention.General
One processor 301 can be microprocessor or any conventional processor etc..The information in conjunction with provided by the embodiment of the present invention
The step of recommended method, can be embodied directly in hardware decoding processor and execute completion, or with the hardware in decoding processor
And software module combination executes completion.Software module can be located in storage medium, which is located at memory, processor
The step of reading the information in memory, completing preceding method in conjunction with its hardware.
In the exemplary embodiment, information recommending apparatus can by one or more application specific integrated circuit (ASIC,
Application Specific Integrated Circuit), DSP, programmable logic device (PLD, Programmable
Logic Device), Complex Programmable Logic Devices (CPLD, Complex Programmable Logic Device), be used for
Execute preceding method.
It in the exemplary embodiment, is the knot for the information recommending apparatus that one embodiment of the invention provides please continue to refer to Fig. 6
Structure schematic diagram, the information recommending apparatus include: receiving module 11, user vector module 13, article vector module 15, enquiry module
18 and recommending module 19.The receiving module 11, the recommendation request of the correspondence target user for receiving client;The user
Vector module 13, for extracting the corresponding different types of user characteristics of the target user, according to the different types of use
Family feature combines the user vector to form the target user;The article vector module 15, for extracting the inhomogeneity of article
The article characteristics of type combine the article vector to form the article according to the different types of article characteristics;The inquiry mould
Block 18 is used for according to the distance between the article vector and the user vector, determining and the user vector to span
Article from the condition that meets;The recommending module 19, for the article based on the condition that meets to the client send pair
The recommendation information answered.
In some embodiments, the user vector module 13 includes coding unit and splicing dimensionality reduction unit, the coding
Unit is used to extract the corresponding different types of user characteristics of the target user from the journal record of the target user, and
Extracted user characteristics are encoded, the coding vector of the user characteristics of each type is obtained;The splicing dimensionality reduction list
Member for the coding vector of different types of user characteristics to be carried out splicing, and drops spliced coding vector
Peacekeeping process of fitting treatment obtains the user vector of the target user.
In some embodiments, the coding unit is also used to the user characteristics of each type are empty from initial code
Between be mapped to newly organized code space, with obtain the user characteristics of each type the newly organized code space it is corresponding it is described encode to
Amount;Wherein, the dimension of the initial code is greater than the dimension of the newly organized code space.
In some embodiments, the coding unit is also used to for the user characteristics of each type being grouped, belong to
The different characteristic group of same physical meaning is using identical Code Mapping Tables from initial code space reflection to newly organized code space.
In some embodiments, the splicing dimensionality reduction unit, the user for being also used to each type of the target user are special
Sign includes different grouping, determines the mean value of the coding vector of the user characteristics in same grouping;It will be each in the different type
The mean value of the coding vector of a grouping is spliced.
In some embodiments, the splicing dimensionality reduction unit is also used to each dimension of spliced coding vector is defeated
Enter the neural network model connected entirely, and user vector of the output par, c dimension as the target user.
In some embodiments, the article vector module is also used to extract the different types of article characteristics of article, will
The mean value of the coding vector of each grouping is spliced in the different type, and is carried out at dimensionality reduction to spliced coding vector
Reason, obtains the article vector of the article.
In some embodiments, the article vector module, be also used to according to the identity of article and article vector it
Between mapping relations, establish the index data base of full dose article.
In some embodiments, the information recommending apparatus further includes sample acquisition module, for obtaining training sample set,
The different type feature and the sample of users that the training sample that the training sample is concentrated includes sample of users are for described
The practical scoring of article;Training module, for the information recommendation model based on training sample training initialization.
In some embodiments, the user vector module is also used to mix the sample with the different types of user characteristics at family
Coding vector carry out splicing and dimensionality reduction and process of fitting treatment, obtain the user vector of sample of users;The article vector
Module is also used to the coding vector of the different types of article characteristics of the article carrying out splicing and dimensionality reduction and intend
Conjunction processing, obtains the article vector of the sample article;The training module is also used to be scored and the reality according to the prediction
Error between the scoring of border, updates the parameter of the information recommendation model, until the loss function of the information recommendation model is received
It holds back.
In some embodiments, the user vector module, the different types of user for being also used to the sample of users are special
Sign includes different grouping, the mean value of the coding vector of the user characteristics in same group is determined, by grouping each in different type
The mean value of coding vector spliced, and dimensionality reduction and process of fitting treatment are carried out to spliced coding vector, obtain the sample
The user vector of user;And the article vector module, the different types of article characteristics for being also used to the article include not
With grouping, determine the mean value of the coding vector of the article characteristics in same group, by the coding of grouping each in different type to
The mean value of amount is spliced, and carries out dimensionality reduction and process of fitting treatment to spliced coding vector, obtain the article of the article to
Amount.
In some embodiments, the sample acquisition module is also used to special from the user of log recording sample drawn user
The article characteristics of sign and the article recommended to the sample of users, by the user characteristics of the sample of users and the article
Article characteristics carry out initial code respectively, mark domain belonging to individual features and the sample of users to be directed to the article
Scoring.
In some embodiments, the sample acquisition module is also used to carry out data cleansing to the log recording, described
Data cleansing includes duplicate removal noise record and/or exception record.
In some embodiments, the coding unit is also used to the text in the user characteristics and the article characteristics
This content carries out modulo operation after being converted to integer according to cryptographic Hash, respectively converts the user characteristics and the article characteristics
For the initial code for specifying data format.
On the other hand, the training method of the information recommendation model provided uses and information recommendation method the embodiment of the present invention
Identical server side is implemented, and for the hardware configuration for implementing the training device of training method of the information recommendation model, asks
It is an optional hardware structural diagram of the training device of information recommendation model provided in an embodiment of the present invention refering to Fig. 7,
The training device of the information recommendation model include: at least one second processor 401, second memory 402, at least one second
Network interface 404 and second user interface 406.Various components in the training device of information recommendation model pass through bus system coupling
It is combined.It should be understood that bus system is for realizing the connection communication between these components.It includes data that bus system, which is removed,
It further include power bus, control bus and status signal bus in addition except bus.It, will in Fig. 4 but for the sake of clear explanation
Various buses are all designated as bus system.
Wherein, second user interface 406 may include display, keyboard, mouse, trace ball, click wheel, key, button,
Touch-sensitive plate or touch screen etc..
It is appreciated that second memory 402 can be volatile memory or nonvolatile memory, may also comprise volatile
Both property and nonvolatile memory.The memory of description of the embodiment of the present invention is intended to include the memory of any suitable type.
Second memory 402 in the embodiment of the present invention is for storing various types of data to support information recommending apparatus
Operation.The example of these data includes: any executable program for operating on information recommending apparatus, such as operating system
And application program;Training sample;Recommend log recording;The index data base etc. of full dose article;Wherein, operating system includes various
System program, for example, ccf layer, core library layer, driving layer etc., hardware based for realizing various basic businesses and processing
Task.Application program may include various application programs, such as media player (Media Player), browser
(Browser) etc., for realizing various applied business.Realize the training method of information recommendation model provided in an embodiment of the present invention
Training device may include in the application.
The method that the embodiments of the present invention disclose can be applied in second processor 401, or by second processor
401 realize.Second processor 401 may be a kind of IC chip, the processing capacity with signal.During realization,
Each step of the above method can pass through the integrated logic circuit of the hardware in second processor 401 or the instruction of software form
It completes.Above-mentioned second processor 401 can be general processor, digital signal processor (DSP, Digital Signal
Processor) either other programmable logic device, discrete gate or transistor logic, discrete hardware components etc..The
Two processors 401 may be implemented or execute disclosed each method, step and logic diagram in the embodiment of the present invention.General
Two processors 401 can be microprocessor or any conventional processor etc..The information in conjunction with provided by the embodiment of the present invention
The step of recommended method, can be embodied directly in hardware decoding processor and execute completion, or with the hardware in decoding processor
And software module combination executes completion.Software module can be located in storage medium, which is located at memory, processor
The step of reading the information in memory, completing preceding method in conjunction with its hardware.
In the exemplary embodiment, the training device of information recommendation model can be by the dedicated integrated electricity of one or more application
Road (ASIC, Application Specific Integrated Circuit), DSP, programmable logic device (PLD,
Programmable Logic Device), Complex Programmable Logic Devices (CPLD, Complex Programmable Logic
Device), for executing preceding method.
It in the exemplary embodiment, is the instruction for the information recommendation model that one embodiment of the invention provides please continue to refer to Fig. 7
Practice the structural schematic diagram of device, the training device of the information recommendation model includes: sample module 21, subscriber-coded module 23, object
Product coding module 25, grading module 27 and convergence module 29.The sample module 21, for obtaining training sample set, the instruction
The different type feature and the sample of users that training sample in white silk sample set includes sample of users are for the article
Practical scoring;The subscriber-coded module 25 uses the sample for passing through the first combination layer of the information recommendation model
The coding vector of the different types of user characteristics at family is combined, and obtains the user vector of sample of users;The article code
Module 25, for passing through the second combination layer of the information recommendation model, by the different types of article characteristics of the article
Coding vector is combined, and obtains the article vector of sample article;Institute's scoring module 27, for passing through the information recommendation mould
Type determines the sample of users to the sample contents according to the vector distance between the user vector and the sample article
The prediction of product is scored;The convergence module 29, for according to prediction scoring and the error update actually to score
The parameter of information recommendation model, until the loss function of the information recommendation model is restrained.Wherein, the first combination layer can be as
One group of coding layer 50, splicing layer 51 and dimensionality reduction layer 52 in information recommendation model shown in Fig. 3.Second combination layer can be such as figure
Another group of coding layer 50, splicing layer 51 and dimensionality reduction layer 52 in information recommendation model shown in 3.
In some embodiments, the subscriber-coded module 23, the different types of user for being also used to mix the sample with family are special
The coding vector of sign carries out splicing and dimensionality reduction and process of fitting treatment, obtains the user vector of sample of users;The article is compiled
Code module 25 is also used to the coding vector of the different types of article characteristics of the article carrying out splicing and dimensionality reduction
And process of fitting treatment, obtain the article vector of the sample article.
In some embodiments, the subscriber-coded module 21 is also used to the different types of user of the sample of users
Feature includes different grouping, determines the mean value of the coding vector of the user characteristics in same group, by each point in different type
The mean value of the coding vector of group is spliced, and carries out dimensionality reduction and process of fitting treatment to spliced coding vector, obtains the sample
The user vector of this user;The article code module 25, be also used to include by the different types of article characteristics of the article
Different groupings determines the mean value of the coding vector of the article characteristics in same group, by the coding of grouping each in different type
The mean value of vector is spliced, and carries out dimensionality reduction and process of fitting treatment to spliced coding vector, obtains the article of the article
Vector.
In some embodiments, the sample module 21, be also used to the user characteristics from log recording sample drawn user,
And the article characteristics for the article recommended to the sample of users, by the object of the user characteristics of the sample of users and the article
Product feature carries out initial code respectively, marks domain belonging to individual features and the sample of users commenting for the article
Point.
In some embodiments, the sample module 21 is also used to carry out data cleansing, the number to the log recording
It include duplicate removal noise record and/or exception record according to cleaning.
The embodiment of the invention also provides a kind of storage medium, the including executable computer program as shown in FIG. 6
One memory 302 or the second memory 402 including executable computer program as shown in Figure 7, above-mentioned computer program can
It is executed by processor, the step of to complete information recommendation method provided by the embodiment of the present invention or executes institute of the embodiment of the present invention
The step of training method of the information recommendation model of offer.Storage medium can be magnetic RAM (FRAM), only
Rdma read (ROM), programmable read only memory (PROM), non-volatile read-only memory (EPROM), electrification can insert programmable
Read memory (EEPROM), flash memory (Flash Memory), magnetic surface storage, CD or compact disc read-only memory (CD-
The memories such as ROM);Be also possible to include one of above-mentioned memory or any combination various equipment, such as computer equipment.
Referring to Fig. 8, the flow diagram of the information recommendation method provided for one embodiment of the invention, can be applied to Fig. 1
Shown in server, will be illustrated in conjunction with following step.
Step 101, the recommendation request of the correspondence target user of client is received.
Client refers to the client that the Information application of information is browsed for user.Receive the correspondence target user of client
Recommendation request, can refer to receive client the identification information for carrying target user recommendation request.The mark of user
Information is to refer to the information of only table sign user identity, such as can be user and logs in the login account of corresponding client, social activity
One of account, user name, counterpart terminal mark etc. or any combination.The correspondence target user's for receiving client pushes away
Request is recommended, can refer to that terminal detects that target user generates recommendation request to the specified operation of client, server receives
The recommendation request of client corresponding with target user.
In some embodiments, the recommendation request of the correspondence target user for receiving client, comprising: detect target
When user's access client, then the recommendation request of the identification information for carrying the target user of client is received.Here,
Access client can refer to that user opens Information application client or refers to and enter after user opens Information application client
To the information page that can receive recommendation information or refer to enter after user opens Information application client it is specified can be with
It browses the information page of information or refers to that user opens Information application client and enters login page and use logs in account
Number successful log etc..
Step 103, the corresponding different types of user characteristics of the target user are extracted, according to the different types of use
Family feature combines the user vector to form the target user.
User characteristics refer to the feature got according to the user identifier of user or attribute, such as from the log recording of user point
The user characteristics of the feature or attribute of precipitation, each type can divide into different groupings according to physical meaning again.It such as, can be with
Including following kind of user characteristics: user base attribute information (including following grouping: age, gender etc.), user are drawn a portrait letter
Cease (including following grouping: classification, label of user of user etc.), (including the grouping below: short-term of user behavior characteristics information
Behavior characteristic information, long-term action characteristic information, such as/long-term classification, label etc. for reading article).From user's
For the user characteristics for extracting target user in log recording, the log recording of the target user for extracting user characteristics
Type, quantity can according to target user's use information apply actual conditions and it is different, using Information application as information
For A, e.g., for the use information application A time relatively long and using more frequently user for, it includes historical behavior
Record it is more, including log recording may include the record of registration information application A, the record of logon information application A, in information
Using browsing article record, the recommendation log recording for obtaining recommendation information by Information application A, user in A in Information application A
Add the record etc. of preference label, then it is described different types of user characteristics are extracted from the log recording of the target user can
With include: from the log recording of multiple and different types of the target user by priority ranking be located at setting range (for example,
Setting range in the sequence of descending since highest priority) in log recording be chosen to be target journaling record, and respectively
Different types of user characteristics are extracted from target journaling record.And the user less for A number of use information application,
Even for the first time using Information application A user for comprising historical behavior record it is less, including log recording can
Can the only record including the record of registration information application A, logon information application A, then described in from the log of target user note
It may include: to remember the log of existing multiple and different types of the target user that different types of user characteristics are extracted in record
It records to select and be recorded as target journaling, and extract choose different types of user spy from target journaling record respectively
Sign.The user vector that user is formed by the corresponding different types of user characteristics of extraction target user, can be to avoid spy
Sparse Problems, and the comprehensive feature that can learn to target user are levied, fully personalized recommendation is thus allowed for.
Step 105, the different types of article characteristics for extracting article combine shape according to the different types of article characteristics
At the article vector of the article.
Article characteristics are the feature got according to article mark or attribute, such as be can be according in recommendation log recording points
The article characteristics or attribute of precipitation.By the different types of article characteristics of article are combined the article of determining article to
Amount, such as the text feature of article, inherent feature and posteriority feature can be combined, it is easy to implement in conjunction with multiple and different
The article characteristics of type, which are merged, obtains accurate article vector expression, can consider different types of article characteristics simultaneously
Between cross influence carry out the attribute of accurate characterization article, obtain the corresponding more comprehensive vector characterization result of each article.
In information recommendation model, the vector for obtaining user and article indicate to be located respectively by independent network
Reason, to can be during information recommendation and extract the article characteristics of full dose article in advance by information recommendation model come really
Determine the corresponding vector expression of full dose article, establish article index data base according to the mapping relations of article and article vector,
By will according to user characteristics determine user vector determine that article vector is mutually indepedent with according to article characteristics, convenient for extend to
The range of matched article vector, so as to facilitate to support the interest extension of user.
Step 107, determining and the user vector according to the distance between the article vector and the user vector
Vector distance meets the article of condition.
Here, the determining vector distance with the user vector meets the article of condition, can refer to according to the target
The distance between user vector and article vector of the finalization of user determine that distance value is less than the article vector of threshold value as full
The article of sufficient condition, or determine article of the relatively small K article of distance value as the condition that meets.
Step 109, corresponding recommendation information is sent to the client based on the article for the condition that meets.
Here, corresponding recommendation information is sent to the client based on the article for the condition that meets, can be institute
Article is stated as selecting to send the first of recommendation information to client and determine range, by determining in range further combined with sieve at the beginning of described
Condition is selected to be screened, using the result after screening as the final recommendation information sent to the client;Alternatively, can also be with
Be according to the quantity of the article for the condition that meets number, when the quantity of the article for the condition that meets is less than threshold value, then
The recommendation information of the article is directly sent to the client.Wherein, for article is article or video, recommendation information
It can be the link address information of article or be chained address and synopsis information of article etc.;It is certain reality for article
For body product, recommendation information can also be the purchase chained address of article or be purchase chained address and the price of article
Information or be purchase chained address and pictorial information of article etc..
In the above embodiment of the present invention, by the recommendation request of acquisition target user, different types of user characteristics are extracted,
The different types of user characteristics of target user are combined to form corresponding user vector, the comprehensive of target user has been arrived in study
Feature;Different types of article characteristics are extracted, different types of article characteristics are combined to form corresponding article vector, are learnt
The comprehensive feature of the article arrived;Article is determined by calculating the distance between article vector and user vector, is effectively reduced
Computation complexity in recommendation process, to realize good recommendation efficiency;By inquiring and the user vector of finalization
Vector distance meets the article of condition, sends corresponding recommendation information to client according to the article.In this way, for different
Target user determines the user characteristics that a certain seed type can be not limited to when corresponding user vector, for arbitrary target user,
Corresponding user vector can be obtained according to its existing different types of user characteristics, for example, can consider target simultaneously
The user behavior characteristics information of user, user different types of user characteristics such as information, user base attribute information of drawing a portrait carry out
Integration, can also be by the intersection shadow between the different types of user characteristics such as the foundation characteristic of user, long-term characteristic, Short-term characteristic
It rings and carries out while considering, so as to obtain more personalized recommendation effect;It, can be with and for different target user
The different types of user characteristics extracted within the scope of different log recordings according to the actual situation, either for new user or
Browse the less user of historical record of article, or for the more old user of historical record of browsing article, it can be real
Now accurate, personalized recommendation, so as to, in different times, can be mentioned for any user or for same user
Rise the accuracy of recommendation information.
In some embodiments, the step 103 extracts the corresponding different types of user characteristics of the target user,
The user vector to form the target user is combined according to the different types of user characteristics, comprising:
Different types of user characteristics corresponding to the target user encode, and the user for obtaining each type is special
The coding vector of sign;
The coding vector of different types of user characteristics is subjected to splicing, and spliced coding vector is dropped
Peacekeeping process of fitting treatment obtains the user vector of the target user.
Wherein, described that extracted user characteristics are encoded, obtain the codings of the user characteristics of each type to
Amount, may include: that the user characteristics of each type are obtained to each class from initial code space reflection to newly organized code space
The user characteristics of type are in the corresponding coding vector of the newly organized code space;Wherein, the dimension of the initial code is greater than institute
State the dimension of newly organized code space.The initial code of the user characteristics can be by the text that will include in extracted user characteristics
This content is converted to by cryptographic Hash, by the user characteristics of each type from initial code space reflection to newly encoded sky
Between, can be the coding lookup Code Mapping Tables according to the user characteristics after conversion, according to feature tag in Code Mapping Tables with
The mapping relations of coding obtain the coding vector of the user characteristics of each type.
Here it is possible to by the coding vector that different types of user characteristics are respectively corresponded to different grouping, then will be different
The coding vector of the user characteristics of type is spliced, to obtain the initialising subscriber vector of the target user, in this way, being directed to
For different user, it can be counted according to different types of user characteristics are obtained respectively the case where the application of its practical use information
Its corresponding user vector is calculated, and is not limited to the user characteristics of one of type.Wherein, it is only limitted to one of type
The mode of user characteristics includes: as being only limitted to determine correspondence according to the user characteristics of user behavior characteristics information this type
Article, that is, content-based recommendation, calculates the correlation between article and article by the content that article includes, according to mesh
The historical record that mark user browses article will recommend the user with the biggish other articles of the article correlation of the historical viewings;
Alternatively, being only limitted to be determined for compliance with the article of condition according to the user characteristics of user base attribute information this type, that is, base
In the recommendation of collaboration, the similitude between user and user is determined according to the historical behavior that user browses article, it will be with wherein one
The article of the biggish other user's concerns of the similitude of user recommends the user;Alternatively, being only limitted to information of drawing a portrait according to user
The user characteristics of this type are determined for compliance with the article of condition, that is, the recommendation based on user's portrait, clear by analysis user
Look at the historical record of article, it is for statistical analysis to the article browsed to determine type, label belonging to these articles,
And by the type of these articles, label mapping to user's portrait, by with the type of corresponding article, label on user's portrait
Identical other articles recommend the user.
In the embodiment of the present invention, all types of user spies is obtained by being encoded to the different types of user characteristics of extraction
The coding vector of sign, and by the coding vector of different types of user characteristics carry out splicing, and to it is spliced encode to
Amount carries out dimensionality reduction and process of fitting treatment, obtains the user vector of the target user, and the user behavior so as to consider user is special
The different types of user characteristics such as reference breath, user's portrait information, user base attribute information are integrated, can also be further
Cross influence between the different feature such as the foundation characteristic of user, long-term characteristic, Short-term characteristic is carried out while being considered, thus
Obtain more personalized recommendation effect;And for different user, different type can be accordingly obtained according to the actual situation
Log recording can be for any user or for same to extract the different types of user characteristics in different range
User is able to ascend the accuracy of recommendation information in different times.
Dimensionality reduction and process of fitting treatment are carried out to spliced coding vector, can be the coding of different types of user characteristics
The user vector that vector is spliced is handled by full Connection Neural Network, more by the study of full Connection Neural Network
The vector of low-dimensional indicates, obtains the user vector that final user's low-dimensional vector table is shown as the target user.This is connected entirely
Neural network correspondence can be the dimensionality reduction layer 52 in information recommendation model shown in Fig. 3.Fig. 9 is please referred to, for the present invention one
The structural schematic diagram of full Connection Neural Network provided by optional embodiment, in full Connection Neural Network, each node
It is connected with each node of preceding layer.It should be noted that carrying out dimensionality reduction and process of fitting treatment to the initial user vector to obtain
The vector expression new to one is also possible to be higher than other neural networks of output layer dimension using input layer dimension, for example, defeated
The middle layer for entering layer and output layer may include one or more hidden layer, such as the knot of multiple convolutional layers and multiple pond layers
Structure reconnects one or more full articulamentums by convolutional layer and pond layer.
In some embodiments, extracted user characteristics are encoded, obtains the user characteristics of each type
Coding vector, comprising:
The user characteristics of each type of the target user are grouped according to physical meaning respectively;
Determine the feature group for belonging to same physical meaning in the group result of different types of user characteristics, and described in determination
Belong to the Code Mapping Tables that the feature group of same physical meaning shares;
Based on the Code Mapping Tables, the user characteristics in the feature group for belonging to same physical meaning are compiled from initial
Code space is mapped to newly organized code space, to obtain the corresponding coding vector;
Wherein, the dimension of the initial code is greater than the dimension of the newly organized code space.
Here, initial code space can be and use such as one-hot for extracted different types of user characteristics
The mode of coding or Hash coding obtains.By the user characteristics of each type from initial code space reflection to newly encoded sky
Between, it can be and be grouped the different types of user characteristics of the target user, the user after grouping is special
Sign determines corresponding vector coding by mapping according to corresponding Code Mapping Tables.Wherein, according to the physical meaning of different characteristic
Feature is divided into different groups, for the feature in identical group, according to feature tag in corresponding Code Mapping Tables and coding
Mapping relations determine the coding vector of character pair.
Code Mapping Tables can refer to that with One-hot code be tabling look-up for index, and the matrix M by establishing n x m defines special
Sign with the linear mapping relation between corresponding more low-dimensional feature vector, so as to by being multiplied to obtain individual features with matrix M
Coding vector, wherein matrix M is Code Mapping Tables (Embedding Table).Wherein, for based on Code Mapping Tables into
Row is searched can be realized with the coding vector correspondence for determining character pair for the coding layer 50 of information recommendation model shown in Fig. 3.
Referring to Fig. 10, being looked for be searched the schematic diagram to determine the coding vector of character pair based on Code Mapping Tables from index
It is encoded to corresponding One-hot Embedding, wherein heavy line directly corresponds to the value of output node, that is, corresponding search is compiled
Code mapping table obtains the coding vector New Embedding coding of feature, to establish the matrix M of Code Mapping Tables as 3x5 matrix
For, Code Mapping Tables are searched according to user characteristics [00010], are closed according to the mapping of feature tag in Code Mapping Tables and coding
System, the coding vector for obtaining user characteristics can indicate as shown in Figure 10.
In the above embodiment of the present invention, by Code Mapping Tables that initial code space reflection is lower to relative dimensions,
Dense space encoder, here processing are properly termed as insertion (embedding) process, and Code Mapping Tables accordingly can be special for index
It levies table (Embedding Table), can solve the sparse problem of feature.
In some embodiments, the user characteristics by each type are from initial code space reflection to newly encoded sky
Between, comprising:
The user characteristics of each type are grouped, belong to the different characteristic group of same physical meaning using identical
Code Mapping Tables are from initial code space reflection to newly organized code space.
Here, feature is divided into according to the physical meaning of different characteristic by different groups, is correspondingly formed multiple feature groups.With with
For the user characteristics of this type of family behavior characteristic information, user behavior characteristics information can be divided into acts and efforts for expediency feature
The different characteristics groups such as information, long-term action characteristic information.For the different characteristic group of same physical meaning, refer to comprising phase jljl
The feature for managing meaning belongs to multiple feature groups simultaneously.Such as the interest characteristics for including in this feature group of acts and efforts for expediency characteristic information,
It is also likely to be simultaneously to belong to this feature group of long-term action characteristic information.Different characteristic group for belonging to same physical meaning is adopted
With identical Code Mapping Tables from initial code space reflection to newly organized code space, refer to the difference spy for same physical meaning
Sign group can share Code Mapping Tables.
It should be noted that the feature tag for same physical meaning can be shared using identical Code Mapping Tables,
By include in user characteristics group long-term set of tags, for short-term set of tags, include in long-term set of tags and short-term set of tags
User characteristics may include the interest tags of same physical meaning, article classification label etc., in this way, can be to same physical meaning
Interest tags, article classification label using identical Code Mapping Tables obtain the coding vector of individual features.Here, for phase
The identical Code Mapping Tables of the shared use of different characteristic group with physical meaning further include the same physical meaning for not same area
Feature, such as the article class target for including in the user characteristics for user behavior characteristics information this type in user domain
Label, and for the set of tags for including in the article characteristics of text feature this type in article domain, it can be to same physical
The article tag feature of meaning obtains the coding vector of individual features using identical Code Mapping Tables.By the way that same physical is contained
The feature tag of justice uses identical Code Mapping Tables, also i.e. by the feature tag of same physical meaning using shared feature to
Amount indicates, can make that the expression of significance of feature is more accurate, reduces feature space, improves recommendation effect.
In some embodiments, the coding vector by different types of user characteristics carries out splicing, comprising:
The grouping of the user characteristics of each type of the corresponding target user, determines user characteristics in each grouping
Coding vector mean value;
The mean value of the coding vector of grouping each in the different type is spliced.
Here, by the feature in same group in such a way that vector is average, available low-dimensional vector table corresponding with group
It reaches.Such as, the user characteristics expression after being grouped to the different types of user characteristics of target user is Ui={ G1Fi1,
G1Fi2,…,G1Fin,…,GmFin, by the low-dimensional vector for obtaining each user characteristics after being mapped based on Code Mapping Tables
It is expressed asFor the user characteristics in same feature group:The low-dimensional vector expression of corresponding group is obtained by the average mode of vector are as follows:Wherein,P is characterized the group number of group.
The mean value of the coding vector by grouping each in the different type is spliced, and can be referred to different groups
Coding vector spliced, obtain the initialising subscriber vector of the target user.Here, for the feature in different groups,
Together by merging features, target user is obtained based on the user vector expression obtained after the fusion of different types of user characteristics.
Such as, the initial of the target user is obtained after being spliced different groups of corresponding low-dimensional vector expression for target user Ui
Changing user vector expression can be with are as follows:
In some embodiments, described that dimensionality reduction and process of fitting treatment are carried out to spliced coding vector, obtain the target
The user vector of user, comprising:
By the neural network model connected entirely, by the spliced coding vector be mapped as the target user to
The vector of quantity space, wherein the dimension of the vector space of the target user is lower than the dimension of the coding vector, and
The vector that mapping obtains is fitted to identical valued space, obtains the user vector of the target user.
Here, it is indicated by the vector that full Connection Neural Network learns more low-dimensional, obtains the final low-dimensional vector of user
Indicate the user vector as the target user.It is complete provided by an optional embodiment of the invention referring to Fig. 9
The structural schematic diagram of Connection Neural Network, in full Connection Neural Network, the node between every satisfactory to both parties articulamentum has Bian Xianglian,
The full Connection Neural Network may include one or more full articulamentum.Wherein, it is reflected by the neural network model connected entirely
The value of coding vector before penetrating may be it is diversified, by the neural network that connects entirely using activation primitive, such as sigmoid
Its value can be fitted to same valued space by function, to obtain the user vector of the target user of more low-dimensional.
It in the above embodiment of the present invention, is indicated by the vector of neural network learning more low-dimensional connected entirely, and by user
It indicates to handle respectively by independent two neural networks connected entirely in information recommendation model with the vector of article, it is mutually deserved
To independent network parameter as a result, to can use information recommendation model and pass through to user characteristics during information recommendation
The corresponding separate network parametric results that are handled obtain the user vector of the final low-dimensional of target user, convenient for real
It is merged now in conjunction with the user characteristics of multiple and different types and obtains accurate user vector and indicate.
In some embodiments, the neural network model by connecting entirely reflects the spliced coding vector
Penetrate the vector of the vector space for the target user, comprising:
The neural network model connected entirely includes at least one active coating;
By activation primitive used in the active coating, by the spliced coding vector to the target user's
Vector space carries out nonlinear mapping;
Mapping result is passed into next active coating in the neural network model to continue non-linearly to reflect
It penetrates, alternatively, mapping result to be exported to the vector of the user's space for the target user.
Wherein, full Connection Neural Network may include one or more active coating, the activation primitive F that active coating uses
(x) line rectification function (Rectified Linear Unit, Relu), threshold function table (Sigmoid), hyperbolic letter can be used
The activation primitives such as number (Tanh).The neural network connected entirely is by using activation primitive, the nerve net that input can be connected entirely
The diversified coding vector of the value of network is fitted to same valued space, to obtain the final target user of more low-dimensional
User vector.
In some embodiments, the step 105 extracts the different types of article characteristics of article, according to the difference
The article characteristics of type combine the article vector to form the article, comprising:
The different types of article characteristics for extracting article, by the mean value of the coding vector of grouping each in the different type
Spliced, and dimensionality reduction and process of fitting treatment are carried out to spliced coding vector, obtains the article vector of the article.
Here, the mean value of the coding vector by grouping each in the different type is spliced, and can be to institute
The article characteristics for stating article are grouped according to different type, by the article characteristics after grouping be mapped to it is corresponding to
Amount coding;The vector coding of the article characteristics in same group is averagely determined to the feature vector of corresponding group by vector, then
Different groups of feature vector is spliced to obtain the corresponding article vector of the article.Wherein, the different type can wrap
Include following at least two: text feature, inherent feature, posteriority feature.
Wherein, different types of article characteristics are handled by information recommendation model to obtain corresponding article vector
When, for each type article characteristics according to the difference of different characteristic physical meaning, different groups can be divided into.With inhomogeneity
The article characteristics of type include text feature, inherent feature and posteriority feature, and article characteristics are divided into set of tags, classification group, theme
Group, set of titles, article ID group, article issuing time group, user's exposure frequency tier group, number of clicks tier group, like time
For the t feature group such as tier group, such as article IjWith a series of article characteristics { Fj1,Fj2,..,Fjn, then it can define object
Product Ij={ Fj1,Fj2,..,Fjn, according to the grouping to article characteristics, article I can be redefinedj={ G1Fj1,G1Fj2,…,
G1Fjn,…,GmFjn, wherein GkFjnIndicate some feature n, G in article j character pair group kmFjnMiddle m indicates article characteristics
The group number of group, k are less than m.Article characteristics after the grouping are encoded by searching for the mode of Code Mapping Tables, will be divided
Each article characteristics after group are mapped to corresponding vector coding, such asFor same
Article characteristics in feature groupThe low-dimensional in corresponding group is obtained by the average mode of vector
Vector expression, whereinWhereinWherein q corresponds to the spy of article
The group number of sign group.For the feature in different characteristic group, feature is spliced, for article Ij, article is obtained after splicing
Initialization vector be expressed asBy by the initial of article
Changing vector indicates to carry out dimensionality reduction and process of fitting treatment by one or more full articulamentum, obtains the more low-dimensional of the finalization of article
Vector table is shown as the article vector of the article.Wherein, the object of the finalization of the user vector and article of the finalization of user
The equal length of product vector.
In some embodiments, the step 107, according to the distance between the article vector and the user vector,
The determining vector distance with the user vector meets before the article of condition, comprising: according to the identity of article and article
Mapping relations between vector establish the index data base of full dose article.
The step 107, comprising: according to each object in the index data base of the user vector and the full dose article
Vector distance between product vector determines the article for meeting condition corresponding with the target user.
Here, it is index with the identity of article, establishes reflecting between identity and article vector comprising article
The index data base of relationship is penetrated, is convenient for during information recommendation, it is convenient based in the index data base comprising full dose article
Article vector calculates separately the vector distance between the user vector of target user and article vector, with determining and target user
Corresponding article.The identity of article is the identification information for referring to the identity of only table sign article, including but not limited to object
Product ID, Item Title and model etc..Wherein, information recommendation model is special based on the article for recommending log recording to extract full dose article
Sign, determines corresponding article vector according to article characteristics, according to the mapping relations between the identity of article and article vector,
The index data base for establishing full dose article can be and be established by the article vector that information recommendation model obtains full dose article offline
Index data base, can also be the corresponding article of the good target user of off-line calculation as recommended candidate, online to target user into
When row information is recommended, corresponding recommendation information can be determined according to the article of the good recommended candidate of off-line calculation.
Vector distance can be to be determined by way of calculating user vector dot product corresponding with each article vector respectively,
During carrying out information recommendation, by the different types of user characteristics of information recommendation model On-line testing target user,
The coding vector of the different types of user characteristics extracted is grouped, is spliced and dimension-reduction treatment, determines target user
The low-dimensional that is merged based on different types of user characteristics user vector, by user vector with pre-establish
Each article vector in the index data base of full dose article calculates vector distance, so that it is determined that satisfaction corresponding with target user
The article of condition.Wherein, the article for meeting condition can refer to relatively nearest K article at a distance from user vector.
In some embodiments, the information recommendation method further include: obtain training sample set, the training sample is concentrated
Training sample include sample of users different type feature and the sample of users be directed to the article practical scoring;
Based on the initial information recommendation model of training sample training.
Wherein, training sample set is obtained, can be according to recommending log recording to be formed includes that the user that same area does not mark is special
Levy the training sample of label and article characteristics label.The recommendation log recording can be according to user's connecing to the information recommended
Positive training sample or negative training sample are divided by situation, optionally, can will expose and selected recommendation log recording is
Positive training sample, exposing unselected recommendation log recording is negative training sample.For Positive training sample, user is characterized to article
Preference corresponding to be scored above preset threshold, in the corresponding Positive training sample, reality of the sample of users to article
Border scoring can be 1;Correspondingly, being directed to negative training sample, characterizes scoring corresponding to preference of the user to article and be lower than
The preset threshold corresponds in the negative training sample, and the sample of users can be 0 to the practical scoring of article.
By forming the training sample of user characteristics label and article characteristics label including not same area mark, so that believing
It ceases in recommended models, the user characteristics and the article characteristics can be handled respectively according to not same area.Secondly, based on just,
The building method of negative training sample, Positive training sample is exposure and selected recommendation log recording, that is, using in Positive training sample
The probability of happening that article is clicked at family is 1, and negative training sample is exposure and unselected recommendation log recording, that is, negative training sample
It is 0 that user, which clicks the probability of happening of article, in this, in this way, by sample of users in each training sample to the preference journey of sample article
The problem of spending is converted to 0-1 classification problem, by the dot product calculated result d (U of user vector and article vectori,Ij) pushed away as information
The input for recommending the loss function of model, exports as 0-1 probability, by the information recommendation model of constantly training iteration adjustment initial
Network parameter obtain what information recommendation model exported until loss function is restrained and the information recommendation model after train
The weight parameter of user vector, article vector and information recommendation model.
In some embodiments, described based on the training sample training information recommendation model, comprising: to mix the sample with
The coding vector of the different types of user characteristics at family carries out splicing and dimensionality reduction and process of fitting treatment, obtains sample of users
User vector;The coding vector of the different types of article characteristics of the article is subjected to splicing and dimensionality reduction and is intended
Conjunction processing, obtains the article vector of the sample article;According to the error between the prediction scoring and the practical scoring, more
The parameter of the new information recommendation model, until the loss function of the information recommendation model is restrained.
Here, in information recommendation model, the coding vector of different types of user characteristics is grouped, is spliced and dimensionality reduction
The user vector of sample of users is obtained, the coding vector of different types of article characteristics is grouped, is spliced and dimension-reduction treatment,
The article vector of sample article is obtained, the sample is determined according to the vector distance between the user vector and the article vector
Preference of this user to the sample article;Instruction is iterated to the information recommendation model based on the training sample
Practice, until the loss function of the information recommendation model is restrained.
Wherein, recommend log recording by obtaining, by forming user characteristics label and article including not same area mark
The training sample of feature tag, the information recommendation model after being trained and being trained to information recommendation model, after the training
Information recommendation model can be used for determining after the recommendation request according to target user and corresponding with the target user meet item
Before the article of part, the different types of user characteristics for extracting the target user generate the user vector of the target user.
The history recommendation information for recommending log recording that can refer to that recommender system is sent to client is formed by log note
Record.Recommendation log recording in recommender system is subjected to feature extraction, extracts user characteristics and article characteristics, respectively by user spy
Article characteristics of seeking peace mark using not same area and accordingly add corresponding user characteristics label and article characteristics label, are formed and are carried
There is the training sample of not the user characteristics label of same area mark and article characteristics label.Since user characteristics and article characteristics are distinguished
Corresponding feature tag is marked using not same area, in this way, in information recommendation model, can according to not same area to user characteristics and
Article characteristics are identified and distinguished between, handled respectively the user characteristics and article characteristics, according to the volume to user characteristics
Code vector is grouped, splicing and dimensionality reduction and process of fitting treatment obtain the user vector of sample of users, according to article characteristics
Coding vector is grouped, splicing and dimensionality reduction and process of fitting treatment obtain the article vector of sample article.Wherein, for user
It is two independent network parameter results with article.
Determine the sample of users to the sample according to the vector distance between the user vector and the article vector
The preference of this article, can refer to by the dot product for calculating user vector and article vector obtain between the two to span
From so that it is determined that the sample of users scores to the prediction of the sample article, prediction scoring can characterize the sample and use
Preference of the family to the sample article.It is with the user vector of sample of usersSample article
Article vector beFor, then the dot product calculation formula of user vector and article vector can be with
Are as follows: d (Ui,Ij)=(x1·y1+x2·y2+…+xn·yn), sample of users is accordingly characterized to sample by the calculated result of dot product
The prediction of article is scored.
The parameter undated parameter of the information recommendation model according to error update refers to, according to prediction scoring and sample of users
For article practical scoring as error signal, the backpropagation in information recommendation model, in each network of backpropagation
In layer, in the way of the solution of various gradients, gradient of the loss function relative to network layer parameter is determined, by the parameter of network layer
It subtracts corresponding gradient and realizes and update.Optionally, the loss function of information recommendation model can be minimized using gradient descent method
Cross entropy loss function.
In the above embodiment of the present invention, information recommendation model is carried out by forming training sample based on recommendation log recording
Training is grouped different types of user characteristics by information recommendation model, splices and dimensionality reduction is fitted to obtain user vector, right
Different types of article characteristics are grouped, splice and dimensionality reduction is fitted to obtain article vector, and user and article are respectively adopted solely
Vertical network processes and can accordingly obtain independent network parameter as a result, after the convergence of information recommendation model training, then pass through
In the present information recommendation process of network parameter fructufy that information recommendation model can independently be handled user and article, root
Obtain the target user's according to the independent different types of user characteristics for extracting target user of the log recording of different range
User vector, can also the independent different types of article characteristics for extracting each article obtain the article vector of article, by basis
The user vector that user characteristics determine determines that article vector is mutually indepedent with according to article characteristics, convenient for extending article to be matched
The range of vector, so as to facilitate to support the interest extension of user;It is also convenient for the day that can be got according to different user
The case where will recording interval and go to get different types of user characteristics and melt the corresponding user vector of joint account, from regardless of
It is for new user or to browse the less user of historical record of article, or the more old use of historical record of browsing article
For family, it is able to achieve accurate, personalized recommendation.
In some embodiments, the coding vector of the different types of user characteristics by the sample of users is spelled
It connects and dimensionality reduction and process of fitting treatment, obtains the user vector of sample of users, comprising: the different types of use of the sample of users
Family feature includes different grouping, determines the mean value of the coding vector of the user characteristics in same group, will be each in different type
The mean value of the coding vector of grouping is spliced, and carries out dimensionality reduction and process of fitting treatment to spliced coding vector, is obtained described
The user vector of sample of users.
Here, can be divided into different for the user characteristics of each type according to the difference of different characteristic physical meaning
Group.It include still user base attribute information, user's portrait information and user behavior characteristics information with different types of user characteristics
For, user base attribute information can be divided into age group, gender group etc.;User draw a portrait information can be divided into Income Classes group,
Interest pattern group etc.;User behavior characteristics information can be divided into long-term set of tags, short-term set of tags etc., can for user characteristics
To be divided into the r feature group such as age group, gender group, Income Classes group, long-term set of tags, short-term set of tags.Such as, user UiHave
Sequence of user feature { Fi1,Fi2,..,Fin, then it can define Ui={ Fi1,Fi2,..,Fin, divide according to user characteristics
Group can redefine U according to each user characteristics corresponding the case where belonging to corresponding groupi={ G1Fi1,G1Fi2,…,
G1Fin,…,GmFin, wherein GkFinIndicate some feature n, G in user i feature group kkFinMiddle m indicates user characteristics group
Group number, k are less than m.
User characteristics after the grouping are encoded by searching for the mode of Code Mapping Tables, it is available each
The low-dimensional vector expression of user characteristics, e.g.,It is special for the user in same feature group
Sign, such as:The low-dimensional vector expression in corresponding group is obtained by the average mode of vector, such as:Wherein,P corresponds to the group number of the feature group of user characteristics.
For the feature in different characteristic group, feature is spliced, for user Ui, user's initialization vector table is obtained after splicing
It is shown as
Spliced feature vector is subjected to dimensionality reduction, the nerve of the dimension of output layer can be greater than using the dimension of input layer
Network indicates progress dimension-reduction treatment to the initialization vector of user, and the vector for obtaining user's finalization indicates.The dimension of input layer
Neural network greater than the dimension of output layer such as can be full Connection Neural Network, may include that one layer or multilayer connect entirely
Layer, will be at the beginning of user for the structure chart for the full Connection Neural Network that can be used in an optional embodiment referring to Fig. 9
Beginningization vector is expressed as As the input of input layer,
Obtain the user vector of finalizationOutput as output layer.Full Connection Neural Network can wrap
Containing one or more active coating, line rectification function (Rectified is can be used in the activation primitive F (x) that active coating uses
Linear Unit, Relu), threshold function table (Sigmoid), the activation primitives such as hyperbolic functions (Tanh).
In some embodiments, the coding vector of the different types of article characteristics by the article spliced,
And dimensionality reduction and process of fitting treatment, obtain the article vector of the sample article, comprising: the different types of article of the article is special
Sign includes different grouping, the mean value of the coding vector of the article characteristics in same group is determined, by grouping each in different type
The mean value of coding vector spliced, and dimensionality reduction and process of fitting treatment are carried out to spliced coding vector, obtain the article
Article vector.
Here, similar to the processing to user characteristics for the processing of article characteristics.For the article characteristics of each type
According to the difference of different characteristic physical meaning, different groups can be divided into.With different types of article characteristics include text feature,
For inherent feature and posteriority feature, text feature is divided into set of tags, classification group, theme group, set of titles etc.;Inherent feature point
For article ID group, article issuing time group etc.;Posteriority feature is divided into user's exposure frequency tier group, number of clicks tier group, point
Praise number tier group etc..For article characteristics, set of tags, classification group, theme group, set of titles, article ID group, article can be divided into
The t feature groups such as issuing time group, user's exposure frequency tier group, number of clicks tier group, like time tier group, such as article
IjWith a series of article characteristics { Fj1,Fj2,..,Fjn, then it can define article Ij={ Fj1,Fj2,..,Fjn, according to object
The grouping of product feature can redefine Ij={ G1Fj1,G1Fj2,…,G1Fjn,…,GmFjn, GkFjnIndicate that article j is corresponding special
Some feature n, G in sign group kmFjnMiddle m indicates that the group number of article characteristics group, k are less than m.
Article characteristics after the grouping are encoded by searching for the mode of Code Mapping Tables, it is available each
The low-dimensional vector expression of article characteristics, such as:It is special for the article in same feature group
Sign, such as:The low-dimensional vector expression in corresponding group is obtained by the average mode of vector,
Such as:Wherein,Q corresponds to the group number of the feature group of article.It is right
Feature in different characteristic group, feature is spliced, for article Ij, the initialization vector table of article is obtained after splicing
It is shown as
Spliced feature vector is subjected to dimensionality reduction, the nerve of the dimension of output layer can be greater than using the dimension of input layer
Network indicates progress dimensionality reduction and process of fitting treatment to the initialization vector of article, and the vector for obtaining article finalization indicates.Input layer
Dimension be greater than the neural network of dimension of output layer and such as can be full Connection Neural Network, may include that one layer or multilayer are complete
Articulamentum is the structure chart for the full Connection Neural Network that can be used in an optional embodiment, by object referring to Fig. 9
Product initialization vector is expressed as As the input of input layer,
Obtain the user vector of finalizationOutput as output layer.Neural network may include one
Or the activation primitives such as Relu, Sigmoid, Tanh can be used in multiple active coatings, corresponding activation primitive G (x).
In some embodiments, the acquisition training sample, comprising: special from the user of log recording sample drawn user
The article characteristics of sign and the article recommended to the sample of users, by the user characteristics of the sample of users and the article
Article characteristics carry out initial code respectively, mark domain belonging to individual features and the sample of users to be directed to the article
Scoring.
Wherein, the user characteristics of the sample drawn user from log recording, and recommend to the sample of users
The article characteristics of the user characteristics of the sample of users and the article are carried out initial code by the article characteristics of article respectively,
It marks domain belonging to individual features and the sample of users to be directed to the scoring of the article, can be according to the recommendation day
Will record extracts user characteristics and article characteristics, and the user characteristics and the article characteristics are carried out code conversion respectively, point
It Tian Jia be not Positive training sample using the user characteristics label and article characteristics label of not same area mark, and to log recording is recommended
Or negative training sample is marked respectively.By the user characteristics label for carrying same area mark and article characteristics label
In the recommendation log recording, exposure and the selected recommendation day for recommending log recording unselected as Positive training sample, exposure
Will record is used as negative training sample, forms training dataset.
It here, include carrying out feature extraction, code conversion and addition according to recommending log recording to form training dataset
Feature tag.Feature extraction includes extracting respectively to the user characteristics and article characteristics recommended in log recording, initial to compile
Code includes being based on cryptographic Hash to the content of text for including to be converted to corresponding feature coding in feature, and addition feature tag includes
User characteristics and article characteristics are marked using not same area, by the way that original user characteristics and article characteristics extraction are gone forward side by side
Row code conversion, it is unified into two feature coding spaces of user and article, so as to realize the unified management to feature
And autocoding, the training sample of the data format by the way that characteristic to be organized into standard are conveniently subsequently used for information recommendation
The training of model.
Wherein, training sample divides into Positive training sample and negative training sample, and Positive training sample is exposure and selected pushes away
Log recording is recommended, e.g., by taking article is video as an example, recommends the video for recommending user in log recording and user is selected and play
Duration is more than that the data of setting time value are Positive training sample;By taking article is article as an example, recommend to recommend use in log recording
The article at family and user selectes and click open data be Positive training sample.And negative training sample is exposure and unselected pushes away
Recommend log recording, e.g., by article be video for, recommend log recording in recommend user video and the unselected broadcasting of user
Or it is negative training sample that playing duration, which is less than the data of setting time value,;By taking article is article as an example, recommend in log recording
Recommend the article of user and user is unselected and to click open data be negative training sample.
In the above embodiment of the present invention, it is directed to the practical scoring of article by sample of users, constructs positive and negative training sample shape
At training dataset, the probability of happening of user's click article is 1 in Positive training sample namely sample of users is directed to the scoring of article
Higher than preset threshold;The probability of happening of user's click article is 0 in negative training sample namely sample of users is directed to the scoring of article
Lower than preset threshold;In this way, sample of users in each training sample is passed through scoring to the problem of preference of sample article
0-1 classification problem is converted to, by the dot product calculated result d (U of user vector and article vectori,Ij) as information recommendation model
The input of loss function exports as 0-1 probability, by constantly training the network parameter in iteration adjustment information recommendation model, directly
Restrained and information recommendation model after train to loss function, it is possible thereby to obtain user that information recommendation model exports to
Amount, the weight parameter of article vector and information recommendation model.
In some embodiments, the acquisition training data sample, further includes: it is clear that data are carried out to the log recording
It washes, the data cleansing includes duplicate removal noise record and/or exception record.
It here, can recommendation day to the feature extracted when obtaining number of training and forming training dataset accordingly
The data of will record carry out data cleansing to be prevented in data with noise record present in duplicate removal daily record data and exception record
Abnormal data or noise data influence final recommendation effect.For example, can will recommend unexposed to use in log recording
The family but recommendation record for being recorded as exposure data is deleted as noise record;It will recommend to belong on the same day in log recording
Request number of times is more than that the recommendation record of ordinary user's request number of times is deleted as exception record.
It, can by carrying out data cleansing to the recommendation log recording for forming training dataset in the above embodiment of the present invention
To ensure the accuracy and validity of training sample, in lift scheme training process the accuracy of the definitive result of weight parameter and
Validity.
In some embodiments, described to carry out the user characteristics of the sample of users and the object article characteristics just respectively
Begin coding, comprising: it is laggard that the content of text in the user characteristics and the article characteristics according to cryptographic Hash is converted to integer
The user characteristics and the article characteristics are converted to the initial code of specified data format by row modulo operation respectively.
Here it is possible to be that original text feature is converted to integer by way of cryptographic Hash, then by the way that this is whole
Several pairs of integer M modulus, the integer for obtaining a 0~M-1 represent this feature.Such as original user feature tag word Fi, by feature
Into Ii after converting, Fi is text, and Ii is integer.By the way that user characteristics and article characteristics are respectively converted into specified data format
Characteristic is automatically converted to the data format of standard by initial code, respectively that the feature of user and article is same to correspondence
Feature coding space in, conveniently realize the unified management and autocoding of feature.
Figure 11 is please referred to, the another aspect of the embodiment of the present invention also provides a kind of training method of information recommendation model, packet
Include following steps:
Step 201, training sample set is obtained, the training sample that the training sample is concentrated includes the inhomogeneity of sample of users
Type feature and the sample of users are directed to the practical scoring of the article;
Here, sample of users, which can refer to, is determined according to recommendation log recording as the training sample of information recommendation model
User.The history recommendation information for recommending log recording that can refer to that information recommendation system is sent to client is formed by log
Record.Recommendation log recording in recommender system is subjected to feature extraction, user characteristics and article characteristics are extracted, respectively by user
Feature and article characteristics are marked using not same area and accordingly add corresponding user characteristics label and article characteristics label, formation are taken
The training sample of user characteristics label and article characteristics label with not same area mark.By special to the user in training sample
Article characteristics of seeking peace are encoded respectively, be can be and are based on breathing out to the content of text for including in relative users feature and article characteristics
Uncommon value is converted to corresponding feature coding, and accordingly the different types of user characteristics based on the sample of users extracted obtain
Corresponding coding vector, and different types of article characteristics accordingly based on the sample article extracted obtain corresponding volume
Code vector.
Step 203, by the first combination layer of the information recommendation model, by the different types of use of the sample of users
The coding vector of family feature is combined, and obtains the user vector of sample of users;
Step 205, by the second combination layer of the information recommendation model, the different types of article of the article is special
The coding vector of sign is combined, and obtains the article vector of sample article;
Here, information recommendation model uses two independent parts to user characteristics and article characteristics, that is, uses first group
It closes layer and the second combination layer is respectively processed, by being trained to information recommendation model, accordingly obtain and user and article
Corresponding independent network weight parameter.
For each type user characteristics according to the difference of different characteristic physical meaning, different groups can be divided into.With
User characteristics can be divided into the r feature group such as age group, gender group, Income Classes group, long-term set of tags, short-term set of tags
Example, according to user UiWith a series of user characteristics { Fi1,Fi2,..,Fin, then it can define Ui={ Fi1,Fi2,..,Fin,
U is redefined according to each user characteristics corresponding the case where belonging to corresponding group according to the grouping to user characteristicsi={ G1Fi1,
G1Fi2,…,G1Fin,…,GmFin, GkFinIndicate some feature n, G in the feature group k of user ikFinMiddle m indicates user characteristics
The group number of group, k are less than m.For the user characteristics after the grouping, vector search (embedding Lookup) can be passed through
Mode is encoded, and the low-dimensional vector for obtaining each user characteristics indicates, such as: For
User characteristics in same feature groupIt is obtained in corresponding group by the average mode of vector
The expression of low-dimensional vector, such as:Wherein,P corresponds to user spy
The group number of the feature group of sign.For the feature in different characteristic group, feature is spliced, for user Ui, after splicing
It is expressed as to user's initialization vector Again with user
Initialization vector indicate input as the first full Connection Neural Network, to the initialization vector progress more low-dimensional of user
Dimensionality reduction process of fitting treatment obtains the user vector of the finalization of sample of users.
It for sample article, is handled using the similar manner with sample of users, for the article characteristics of each type
According to the difference of different characteristic physical meaning, different groups can be divided into.It is divided into set of tags, classification group, theme with article characteristics
Group, set of titles, article ID group, article issuing time group, user's exposure frequency tier group, number of clicks tier group, like time
For the t feature group such as tier group, according to article IjWith a series of article characteristics { Fj1,Fj2,..,Fjn, then it can define
Article Ij={ Fj1,Fj2,..,Fjn, according to the grouping to article characteristics, I can be redefinedj={ G1Fj1,G1Fj2,…,
G1Fjn,…,GmFjn, GkFjnIndicate some feature n, G in article j character pair group kmFjnMiddle m indicates the group of article characteristics group
Number, k are less than m.For the article characteristics after the grouping, can by way of vector search (embedding Lookup) into
Row coding obtains the low-dimensional vector expression of each article characteristics, such as: For same spy
Article characteristics in sign group, such as:It is obtained in corresponding group by the average mode of vector
The expression of low-dimensional vector, such as:Wherein,Q corresponds to the spy of article
The group number of sign group.For the feature in different characteristic group, feature is spliced, for article Ij, article is obtained after splicing
Initialization vector be expressed as Again with the initialization of article
Vector table is shown as the input of the second full Connection Neural Network, and the dimensionality reduction for carrying out more low-dimensional to the initialization vector of user is fitted
Processing, obtains the article vector of the finalization of sample article.
Wherein, the activation primitive G (x) of the full connection nerve of the activation primitive F (x) of the first full Connection Neural Network and second can
To use Relu, the activation primitives such as Sigmoid, Tanh respectively.
Step 207, by the information recommendation model according between the user vector and the sample article to
Span scores to the prediction of the sample article from the determination sample of users;
Step 209, according to the ginseng of the prediction scoring and information recommendation model described in the error update actually to score
Number, until the loss function of the information recommendation model is restrained.
Here, determine the sample of users to institute according to the vector distance between the user vector and the article vector
The prediction scoring for stating sample article, can refer to by the dot product for calculating user vector and article vector obtain between the two to
Span is from according to the vector distance so that it is determined that prediction scoring of the sample of users to the sample article, passes through prediction
Preference of the scoring characterization sample of users to article.It is with the user vector of sample of users
The article vector of sample article is For, then the dot product of user vector and article vector calculates public
Formula can be with are as follows: d (Ui,Ij)=(x1·y1+x2·y2+…+xn·yn), sample of users is accordingly characterized according to the calculated result of dot product
To the preference of sample article.
Referred to according to the prediction scoring and the parameter of information recommendation model described in the error update actually to score, root
It is predicted that scoring and sample of users for article practical scoring as error signal, reversely passed in information recommendation model
It broadcasts, in each network layer of backpropagation, in the way of the solution of various gradients, determines that loss function is joined relative to network layer
The parameter of network layer is subtracted corresponding gradient and realizes update by several gradients.Here, the loss function of information recommendation model can be with
Cross entropy loss function is minimized using gradient descent method, for the building method of positive and negative training sample, Positive training sample is to expose
Light and selected recommendation log recording, that is, it is 1 that user, which clicks the probability of happening of article, in Positive training sample, negative training sample
For exposure and unselected recommendation log recording, that is, it is 0 that user, which clicks the probability of happening of article, in negative training sample, in this way,
Sample of users in each training sample is converted into 0-1 classification problem to the computational problem of the preference of sample article, will be used
The dot product calculated result d (U of family vector sum article vectori,Ij) as information recommendation model loss function input, export and be
0-1 probability, the network parameter by way of constantly training iteration in adjustment information recommended models, until loss function convergence and
Information recommendation model after being trained, and the user vector of the information recommendation model output after being trained, article vector with
And the weight parameter of information recommendation model.
In the above embodiment of the present invention, by by information recommendation model to the different types of user characteristics of sample of users into
Row combination obtains user vector, is combined to obtain article vector to the different types of article characteristics of sample article, to user
Independent network parameter can be accordingly obtained as a result, working as information recommendation model training extremely using independent network processes with article
After convergence, then it can be mentioned by information recommendation model in information recommendation process according to the log recording of different range is independent
It takes the different types of user characteristics of target user to obtain the user vector of the target user, independent can also extract each object
The different types of article characteristics of product obtain the article vector of article, by the user vector determined according to user characteristics and according to object
Product feature determines that article vector is mutually indepedent, convenient for extending the range of article vector to be matched, so as to facilitate to support
The interest of user extends;The case where being also convenient for the log recording range that can be got according to different user goes to get inhomogeneity
The user characteristics of type calculate corresponding user vector, use either less for the historical record of new user or browsing article
For family, or the more old user of historical record of browsing article, it is able to achieve accurate, personalized recommendation.
In some embodiments, the coding vector of the different types of user characteristics by the sample of users carries out group
It closes, obtains the user vector of sample of users, comprising: the coding vector for mixing the sample with the different types of user characteristics at family is spelled
Processing and dimensionality reduction and process of fitting treatment are connect, the user vector of sample of users is obtained;
The coding vector of the different types of article characteristics by the article is combined, and obtains the object of sample article
Product vector, comprising: the coding vector of the different types of article characteristics of the article is subjected to splicing and dimensionality reduction and is intended
Conjunction processing, obtains the article vector of the sample article.
By being encoded to obtain the coding vector of all types of user characteristics to the different types of user characteristics of extraction, and
The coding vector of different types of user characteristics is subjected to splicing, and dimensionality reduction and fitting are carried out to spliced coding vector
Processing obtains the user vector of the target user, so as to consider that it is whole that a plurality of types of user characteristics of target user carry out
It closes.Correspondingly, by being encoded to obtain the coding vector of all types of article characteristics to the different types of article characteristics of extraction,
And the coding vector of different types of article characteristics is subjected to splicing, and dimensionality reduction is carried out to spliced coding vector and is intended
Conjunction handles to obtain the article vector of article, so as to consider more fully article characteristics.To extract user from log recording
Different types of user characteristics for, for different user, can also accordingly obtain different type according to the actual situation
Log recording extract the different types of user characteristics in different range, can be for any user or for same
User is able to ascend the accuracy of recommendation information in different times.To extract article not from recommendation log recording
For the article characteristics of same type, the article characteristics of full dose article can be extracted and determine corresponding article vector, can establish
The article vector library of full dose article.
In some embodiments, the coding vector by different types of user characteristics carries out splicing and drop
Peacekeeping process of fitting treatment obtains the user vector of sample of users, comprising: the different types of user characteristics of the sample of users include
Different groupings determines the mean value of the coding vector of the user characteristics in same group, by the coding of grouping each in different type
The mean value of vector is spliced, and carries out dimensionality reduction and process of fitting treatment to spliced coding vector, obtains the sample of users
User vector;The coding vector by different types of article characteristics carries out splicing and dimension-reduction treatment, obtains sample
The article vector of article, comprising: the different types of article characteristics of the article include different groupings, are determined in same group
The mean value of the coding vector of article characteristics splices the mean value of the coding vector of grouping each in different type, and to spelling
Coding vector after connecing carries out dimensionality reduction and process of fitting treatment, obtains the article vector of the article.
It, can be by the different types of user characteristics of user by the way that different types of user characteristics are grouped and are spliced
It is merged to determine the corresponding user vector of user, e.g., by user base attribute information, user behavior characteristics information and use
The different types of user characteristics such as family vector information are all stitched together, at the same consider the foundation characteristic of user, long-term characteristic,
Short-term characteristic etc. is integrated to obtain user vector, to obtain user vector by information recommendation model, realizes user's
Cross influence between different types of user characteristics accounts for simultaneously, obtains more personalized recommendation effect;And pass through
Different types of article characteristics are grouped and are spliced, the different types of article characteristics of article can be merged with true
The corresponding article vector of earnest product, it is e.g., the different types of article such as the text feature of article, inherent feature and posteriority feature is special
Text feature and the example aspects etc. that sign is all stitched together, while considering article are integrated to obtain article vector, thus
Article vector is obtained by information recommendation model, is realized the cross influence between the different types of article characteristics of article simultaneously
It accounts for, in this way, article and user to have been carried out to the semantic matches of more depth, improves recommendation effect.
In some embodiments, the coding vector of the different types of user characteristics by the sample of users carries out group
It closes, obtains the user vector of sample of users, comprising: adjust using weight of the attention mechanism to different types of user characteristics
It is whole, the different types of user characteristics are combined to obtain the user vector of sample of users according to respective weights.It is described to incite somebody to action
The coding vector of the different types of article characteristics of the article is combined, and obtains the article vector of sample article, comprising: adopt
It is adjusted with weight of the attention mechanism to different types of article characteristics, to the different types of article characteristics according to phase
Weight is answered to be combined to obtain the article vector of sample article.
Wherein, attention mechanism refers to that model is different to the attention of user's difference behavior when prediction.
Figure 12 is please referred to, the principle to be adjusted using attention (Attention) mechanism to the weight of different types of feature is shown
It is intended to, can be and the feature in model training is considered as by a series of feature and weighted value<Key, Value>data are to structure
At, it is at this time the influence to the result for determining user and item associations to some element Query in setting the goal, such as Query,
By calculating the similitude or correlation of Query and each Key, the corresponding weight coefficient Value of each Key is obtained, then
Summation is weighted to get the Value numerical value for having arrived final Key to Value.It is compiled according to different types of user characteristics
Before code, splicing, dimensionality reduction and fitting obtain corresponding user vector, different types of user characteristics are determined using attention mechanism
Respective weights, the respective weights can be what the article characteristics once bought based on calculating user characteristics and user were determined.For example,
The article characteristics that product/user characteristics and user between the article characteristics that user characteristics and user once bought once were bought
The connection for the article characteristics that difference, user characteristics and user once bought (can be mapped to article/user by full articulamentum
Space).
By taking user vector as an example, it is assumed that user vector Vu, candidate item article vector be Va, the i-th row of user U
For feature vector be Vi, the different types of feature is combined to obtain most according to respective weights based on attention mechanism
Whole vector expression can be such that
Wherein, g (Vi,Va) be denoted as user U i-th behavior feature vector be Vi corresponding to weight.
By the way that attention mechanism is added, different types of feature can be directed to according to it to determining user and item associations
Result influence and form different weights, promoted by training after model carry out information recommendation accuracy.
It can be further understood from order to be provided for the embodiments of the invention the training method of information recommendation model,
Figure 13 is please referred to, the training method of information recommendation model is further illustrated below with reference to a schematical embodiment, is asked
In conjunction with refering to fig. 14, which mainly includes that feature extraction 61, training data generation 62 and model training 63 3 are main
Link includes the following steps:
Step S21 carries out feature extraction according to the recommendation log of information recommendation system, by the user characteristics being drawn into and object
Product feature is converted by cryptographic Hash, unified into two feature coding spaces of user and article;In this way, realizing that feature is taken out
Take link.
The log recording extracted is carried out data cleansing, is based on not according to user characteristics and article characteristics by step S22
The log recording of same area mark generates training sample set;By carry out data cleansing, can in duplicate removal log recording noise or
Abnormal data, it is ensured that the validity of training sample;In this way, realizing that training data generates link.
Training sample set, is inputted initial information recommendation model by step S23, and information recommendation model is independent by two
Network portion is grouped processing to user characteristics and article characteristics respectively, and the feature vector in group is averaged, the feature between group
Vector is spliced, and identical group of physical meaning of feature space of different groups and feature is shared, is learnt each sample respectively and is used
The expression of the vector of family and sample article;
Figure 15 is please referred to, is the framework of the training of the vector expression of each user of information recommendation model learning and article
Schematic diagram, wherein user characteristics F1, user characteristics F2 ... user characteristics Fm respectively indicates different user characteristics groups, passes through
Vector search (embedding lookup) operation will be passed through in different user feature group to encode, available each user
The low-dimensional vector of feature group indicates;Article characteristics F1, article characteristics F2 ... it is special that article characteristics Fm respectively indicates different article
Sign group being encoded by will pass through vector search (embedding lookup) operation in different article characteristics groups, can be obtained
Low-dimensional vector to each article characteristics group indicates.Wherein, it is compiled by vector search (embedding lookup) operation
Code can be realized in the following way.In an alternative embodiment, it encoded, spelled according to different types of user characteristics
It connects, before dimensionality reduction and fitting obtain corresponding user vector, further includes that different types of user characteristics are determined based on attention mechanism
Respective weights, which can be based on calculating what the article characteristics that user characteristics and user once bought determined.For example, with
The difference for the article characteristics that product/user characteristics and user between the article characteristics that family feature and user once bought once were bought
The connection for the article characteristics that value, user characteristics and user once bought (can be mapped to article/user sky by full articulamentum
Between).
Figure 16 is please referred to, for the user vector { F11, F12...F1n } in same user characteristics group F1, by looking into
Corresponding Code Mapping Tables (Embedding Table 1) are looked for, the average mode determination of vector and corresponding user characteristics group are passed through
Corresponding low-dimensional vector indicates F1;For the user vector { F21, F22...F2n } in same user characteristics group F2, by searching for
Corresponding Code Mapping Tables (Embedding Table 2) pass through the average mode determination of vector and corresponding user characteristics group pair
The low-dimensional vector answered indicates F2;And so on, for the user vector { Fm1, Fm2...Fmn } in same user characteristics group Fm,
By searching for corresponding Code Mapping Tables (Embedding Table m), pass through the average mode determination of vector and corresponding user
The corresponding low-dimensional vector of feature group indicates Fm.
Feature vector between group carries out splicing and refers to splice the corresponding low-dimensional vector expression of different user feature group,
As described in Figure 12, by user characteristics group F1, user characteristics group F2 ... the corresponding low-dimensional vector of user characteristics group Fm indicates
F1, F2 ... Fm is spliced, and the corresponding user vector expression { F1, F2...Fm } of user is obtained.
Wherein, the mode phase that two independent network portions of information recommendation model are respectively processed user and article
Seemingly, the framework of two independent network portions is identical, accordingly obtains independent network parameter result respectively by training.It is optional
, each network portion may include realize to the feature after grouping carry out it is average in group and feature between group is spliced the
One layer and the output result of first layer is obtained to realize the full Connection Neural Network of dimension-reduction treatment.Information recommendation model further includes
With the classification layer S (X) (softmax) of the output connection of two independent network portions.
The vector of step S24, the sample of users handled respectively according to information recommendation model and sample article is expressed, meter
User is calculated to the preference of article, the corresponding obtained user of each training sample is accordingly converted to the preference of article
Two classification problems, continuous repetitive exercise adjust network weight parameter, until information recommendation model is restrained.In this way, step S23 and
S24 implementation model trains link.
Wherein, two independent network portions of information recommendation model can respectively the object after the grouping to not same area it is special
By way of sign is encoded to obtain coding vector, and averagely feature is spliced between group feature progress vector in organizing, obtain
To the low-dimensional feature vector of the initialization of corresponding object.Here, the characteristics of objects of same area does not refer respectively to user characteristics and article
Feature, softmax layers can calculate the object low-dimensional feature vector of the initialization of two independent network portion output,
The matching degree between corresponding object is obtained, namely obtains the matching degree between user and article.Information recommendation model passes through
Log recording will be recommended to be trained as training sample, according to the actual match recommended in log recording between user and article
The positive and negative training sample that degree is distinguished, so as to which the matching problem between user and article is converted to two classification problems,
By continuous repetitive exercise loss function L (X) is restrained, the information recommendation model after being trained.
Two independent network portions of information recommendation model accordingly obtain independent network parameter knot by training respectively
Fruit can pass through the different type of the extraction target user of the information recommendation Model Independent after training when carrying out information recommendation
User characteristics merged and export the user vector of corresponding low-dimensional.The information recommendation model after training can also be passed through
The independent different types of article characteristics for extracting article are merged and are exported the article vector of corresponding low-dimensional, are established complete
Measure the index data base of article, convenient between the subsequent user vector based on the target user being calculated and article vector to
Span is from determining qualified article corresponding with target user.
It can be further understood from, please refer in order to be provided for the embodiments of the invention information recommendation method
Figure 17 is further illustrated information recommendation method below with reference to a schematical embodiment, please refers to Figure 13, this is pushed away
The method of recommending may include 65 two key links of on-line prediction 64 and/or K-NN search, include the following steps:
Step S25 constructs the index data base of article based on the information recommendation model after training;It is obtained based on step S23
Training after information recommendation model, using the feature of full dose article determine full dose article low-dimensional vector indicate, and according to
Mapping relations between article ID and article vector are stored, and the index data base of full dose article is constructed;
When receiving the recommendation request of target user, it is special to extract different types of user from log recording by step S26
Sign is encoded after being grouped by the information recommendation model after training according to the different types of user characteristics extracted, is spelled
It connects and dimensionality reduction and fitting, obtains the corresponding low-dimensional user vector of target user;
Different types of user characteristics may include: the primary attribute information of user, such as gender, age, city, income water
The interest classification of flat, Social Grading etc., interest attribute information, such as user, label, theme, special features such as user's serial number etc.,
Acts and efforts for expediency characteristic information, such as classification, the label feature of the K piece article read recently of user.It should be noted that can root
The user characteristics in different range are accordingly extracted according to the actual conditions of different user, e.g., for without the new of recommendation log recording
For user, can according to its corresponding registration log recording, log in log recording etc. and accordingly extract different types of user
Feature, and for the old user for having more recommendation log recording, article record, additive can be browsed in the recent period in conjunction with it
Product interest tags record recommends log recording etc. and accordingly extracts different types of user characteristics.In the step 26, it can be
The recommendation request of real-time reception target user is combined by the different types of user characteristics of information recommendation model extraction after training
User vector is obtained, to pass through information recommendation model realization on-line prediction link.
Wherein, it carries out encoding can be after being grouped the different types of user characteristics extracted passing through embedding
The mode of lookup operation is realized, by the way that original user feature based on the feature coding obtained after Hash translation, is searched and corresponded to
Code Mapping Tables Embedding Table obtain each user characteristics low-dimensional vector indicate.Wherein, for belonging to phase jljl
The different characteristic group for managing meaning, can share Code Mapping Tables, in this way, expression of significance is more in the low-dimensional combination for considering feature
To be accurate, so as to reduce the training speed of feature space, lift scheme, the accuracy of model learning, Jin Erke are also improved
To improve recommendation effect.
Step S27, according to the article of each article in the index data base of the corresponding user vector of target user and article
Vector distance between vector determines article corresponding with target user.
Here it is possible to the dot product for calculating user vector and article vector obtains vector distance between the two, so that it is determined that
The corresponding article of target user.The user vector of target user isThe article vector of article isFor, then the dot product calculation formula of user vector and article vector are as follows: d (Ui,Ij)=(x1·y1
+x2·y2+…+xn·yn), wherein m=n.Wherein, it establishes the index data base of article and determines the user vector of target user
It can also be predefined by the information recommendation model after training, it then can be directly according to precalculated when being recommended
As a result accordingly recommended, to pass through information recommendation model realization K-NN search link.
Step 28, it is screened according to the article based on the screening conditions of setting, determines corresponding recommendation information to institute
State target user's transmission.
Here, it according to the distance between the user vector of target user and article vector, will can determine and user vector
Between the relatively small K article of vector distance as corresponding article.It is formed according to the corresponding article to be recommended
Article collection, then pass sequentially through thick row, essence row and mixing and article is further screened, determine final recommendation information to target
User sends.
In the above embodiment of the present invention, on the one hand, information recommendation model is by user base attribute, attribute of drawing a portrait, short-term row
It is all stitched together for attribute etc., considers the foundation characteristic of user, long-term characteristic, Short-term characteristic is integrated, and is passed through
Deep neural network generates user vector and takes into account the cross influence between user's different characteristic, obtains more personalized
Recommendation effect;On the other hand, information recommendation model is by the attributive character (such as label, classification, topic) of article side, ID feature etc.
Feature is all stitched together, and has both considered the text feature of article, it is contemplated that the example aspects of article, by article and user
The semantic matches of more depth are carried out, to promote recommendation effect;It is based on the matched recommended method of deep semantic in this way, realizing,
By the primary attribute information of user, user draw a portrait (classification, label) information, user's acts and efforts for expediency feature (read article classification,
Label) information etc. and the text information of article, label information etc. matched, and it is low by carrying out user characteristics and article characteristics
Dimensional vector indicates, the similarity between user and article is measured by the dot product calculated between two vectors, for each user
Taking indicates that the highest M article of similarity determines recommendation information with user's low-dimensional vector.To well solve user's base
Comprehensive use and the cross influence between different characteristic of the information such as plinth attribute, user's portrait, acts and efforts for expediency, more personalized
Article recommend user.In another aspect, information recommendation model is when considering the combination of the low-dimensional of feature of user or article, it will not
The identical meanings feature of same area is indicated using shared vector, so that expression of significance is more acurrate, is reduced feature space, is improved
Model training speed improves recommendation effect to improve the accuracy of model learning.
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
Those familiar with the art in the technical scope disclosed by the present invention, can easily think of the change or the replacement, and should all contain
Lid is within protection scope of the present invention.Therefore, protection scope of the present invention should be based on the protection scope of the described claims.
Claims (15)
1. a kind of information recommendation method characterized by comprising
Receive the recommendation request of the correspondence target user of client;
The corresponding different types of user characteristics of the target user are extracted, combine shape according to the different types of user characteristics
At the user vector of the target user;
The different types of article characteristics for extracting article, combine to form the article according to the different types of article characteristics
Article vector;
According to the distance between the article vector and the user vector, the determining vector distance with the user vector meets
The article of condition;
Corresponding recommendation information is sent to the client based on the article for the condition that meets.
2. the method as described in claim 1, which is characterized in that described to extract the corresponding different types of use of the target user
Family feature combines the user vector to form the target user according to the different types of user characteristics, comprising:
Different types of user characteristics corresponding to the target user encode, and obtain the user characteristics of each type
Coding vector;
The coding vectors of different types of user characteristics is subjected to splicing, and to spliced coding vector carry out dimensionality reduction and
Process of fitting treatment obtains the user vector of the target user.
3. method according to claim 2, which is characterized in that it is described that extracted user characteristics are encoded, it obtains each
The coding vector of the user characteristics of the type, comprising:
The user characteristics of each type of the target user are grouped according to physical meaning respectively;
It determines the feature group for belonging to same physical meaning in the group result of different types of user characteristics, and belongs to described in determination
The Code Mapping Tables that the feature group of same physical meaning shares;
It is based on the Code Mapping Tables, the user characteristics in the feature group for belonging to same physical meaning are empty from initial code
Between be mapped to newly organized code space, to obtain the corresponding coding vector;
Wherein, the dimension of the initial code is greater than the dimension of the newly organized code space.
4. method according to claim 2, which is characterized in that the coding vector by different types of user characteristics carries out
Splicing, comprising:
The grouping of the user characteristics of each type of the corresponding target user, determines the volume of user characteristics in each grouping
The mean value of code vector;
The mean value of the coding vector of grouping each in the different type is spliced.
5. method according to claim 2, which is characterized in that described to be carried out at dimensionality reduction and fitting to spliced coding vector
Reason, obtains the user vector of the target user, comprising:
By the neural network model connected entirely, the vector that the spliced coding vector is mapped as the target user is empty
Between vector, wherein the dimension of the vector space of the target user be lower than the coding vector dimension, and
The vector that mapping obtains is fitted to identical valued space, obtains the user vector of the target user.
6. method as claimed in claim 5, which is characterized in that the neural network model by connecting entirely, by the spelling
Coding vector after connecing is mapped as the vector of the vector space of the target user, comprising:
The neural network model connected entirely includes at least one active coating;
By activation primitive used in the active coating, by the spliced coding vector to the vector of the target user
Space carries out nonlinear mapping;
Mapping result is passed into next active coating in the neural network model to continue non-linearly to map, or
Mapping result is exported the vector of the user's space for the target user by person.
7. the method as described in claim 1, which is characterized in that the different types of article characteristics for extracting article, according to
The different types of article characteristics combine the article vector to form the article, comprising:
The different types of article characteristics for extracting article carry out the mean value of the coding vector of grouping each in the different type
Splicing, and dimensionality reduction and process of fitting treatment are carried out to spliced coding vector, obtain the article vector of the article.
8. method as described in any one of claim 1 to 7, which is characterized in that described according to the article vector and the use
The distance between family vector, the determining vector distance with the user vector meet before the article of condition, comprising:
According to the mapping relations between the identity of article and article vector, the index data base of full dose article is established.
9. a kind of training method of information recommendation model characterized by comprising
Obtain training sample set, the training sample that the training sample is concentrated include sample of users different type feature and
The sample of users is directed to the practical scoring of the article;
By the first combination layer of the information recommendation model, by the coding of the different types of user characteristics of the sample of users
Vector is combined, and obtains the user vector of sample of users;
By the second combination layer of the information recommendation model, by the coding vector of the different types of article characteristics of the article
It is combined, obtains the article vector of sample article;
Institute is determined according to the vector distance between the user vector and the sample article by the information recommendation model
Sample of users is stated to score to the prediction of the sample article;
According to the parameter of the prediction scoring and information recommendation model described in the error update actually to score, until the letter
Cease the loss function convergence of recommended models.
10. method as claimed in claim 9, which is characterized in that the different types of user by the sample of users is special
The coding vector of sign is combined, comprising:
The coding vector for mixing the sample with the different types of user characteristics at family carries out splicing and dimensionality reduction and process of fitting treatment,
Obtain the user vector of sample of users;
The coding vector of the different types of article characteristics by the article is combined, comprising:
The coding vector of the different types of article characteristics of the article is subjected to splicing and dimensionality reduction and process of fitting treatment,
Obtain the article vector of the sample article.
11. method as claimed in claim 10, which is characterized in that the different types of user by the sample of users is special
The coding vector of sign carries out splicing and dimensionality reduction and process of fitting treatment, obtains the user vector of sample of users, comprising:
The different types of user characteristics of the sample of users include different groupings, determine the volume of the user characteristics in same group
The mean value of code vector splices the mean value of the coding vector of grouping each in different type, and to it is spliced encode to
Amount carries out dimensionality reduction and process of fitting treatment, obtains the user vector of the sample of users;
The coding vector of the different types of article characteristics by the article carries out at splicing and dimensionality reduction and fitting
Reason, obtains the article vector of the sample article, comprising:
The different types of article characteristics of the article include different groupings, determine the codings of the article characteristics in same group to
The mean value of amount splices the mean value of the coding vector of grouping each in different type, and to spliced coding vector into
Row dimensionality reduction and process of fitting treatment obtain the article vector of the article.
12. a kind of information recommending apparatus characterized by comprising
Receiving module, the recommendation request of the correspondence target user for receiving client;
User vector module, for extracting the corresponding different types of user characteristics of the target user, according to the inhomogeneity
The user characteristics of type combine the user vector to form the target user;
Article vector module, for extracting the different types of article characteristics of article, according to the different types of article characteristics
Combination forms the article vector of the article;
Enquiry module, for according to the distance between the article vector and the user vector, the determining and user vector
Vector distance meet the article of condition;
Recommending module sends corresponding recommendation information to the client for the article based on the condition that meets.
13. a kind of training device of information recommendation model characterized by comprising
Sample module, for obtaining training sample set, the training sample that the training sample is concentrated includes the difference of sample of users
Type feature and the sample of users are directed to the practical scoring of the article;
Subscriber-coded module, for passing through the first combination layer of the information recommendation model, by the inhomogeneity of the sample of users
The coding vector of the user characteristics of type is combined, and obtains the user vector of sample of users;
Article code module, for passing through the second combination layer of the information recommendation model, by the different types of of the article
The coding vector of article characteristics is combined, and obtains the article vector of sample article;
Grading module, for by the information recommendation model according between the user vector and the sample article to
Span scores to the prediction of the sample article from the determination sample of users;
Module is restrained, for according to the ginseng of the prediction scoring and information recommendation model described in the error update actually to score
Number, until the loss function of the information recommendation model is restrained.
14. a kind of computer equipment characterized by comprising processor and by store can run on a processor based on
The memory of calculation machine program;
Wherein, the processor is for realizing claim 1 to 8 described in any item information when running the computer program
The training method of information recommendation model described in any one of recommended method or realization claim 9 to 11.
15. a kind of storage medium, is stored thereon with computer program, which is characterized in that the computer program is executed by processor
Described in any one of the described in any item information recommendation methods of Shi Shixian claim 1 to 8 or realization claim 9 to 11
Information recommendation model training method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910326702.1A CN110162700A (en) | 2019-04-23 | 2019-04-23 | The training method of information recommendation and model, device, equipment and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910326702.1A CN110162700A (en) | 2019-04-23 | 2019-04-23 | The training method of information recommendation and model, device, equipment and storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110162700A true CN110162700A (en) | 2019-08-23 |
Family
ID=67638677
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910326702.1A Pending CN110162700A (en) | 2019-04-23 | 2019-04-23 | The training method of information recommendation and model, device, equipment and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110162700A (en) |
Cited By (46)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110704739A (en) * | 2019-09-30 | 2020-01-17 | 汉海信息技术(上海)有限公司 | Resource recommendation method and device and computer storage medium |
CN110766456A (en) * | 2019-10-16 | 2020-02-07 | 无线生活(杭州)信息科技有限公司 | Commodity recommendation method and device |
CN110781321A (en) * | 2019-08-28 | 2020-02-11 | 腾讯科技(深圳)有限公司 | Multimedia content recommendation method and device |
CN110928990A (en) * | 2019-10-31 | 2020-03-27 | 南方电网调峰调频发电有限公司 | Method special for recommending standing book data of power equipment based on user portrait |
CN111008321A (en) * | 2019-11-18 | 2020-04-14 | 广东技术师范大学 | Recommendation method and device based on logistic regression, computing equipment and readable storage medium |
CN111159564A (en) * | 2019-12-31 | 2020-05-15 | 联想(北京)有限公司 | Information recommendation method and device, storage medium and computer equipment |
CN111242318A (en) * | 2020-01-13 | 2020-06-05 | 拉扎斯网络科技(上海)有限公司 | Business model training method and device based on heterogeneous feature library |
CN111259222A (en) * | 2020-01-22 | 2020-06-09 | 北京百度网讯科技有限公司 | Article recommendation method, system, electronic device and storage medium |
CN111369315A (en) * | 2020-02-27 | 2020-07-03 | 拉扎斯网络科技(上海)有限公司 | Resource object recommendation method and device, and data prediction model training method and device |
CN111368034A (en) * | 2020-03-02 | 2020-07-03 | 复旦大学 | Bidirectional semantic feature matching method and supply content recommendation device |
CN111400613A (en) * | 2020-03-17 | 2020-07-10 | 苏宁金融科技(南京)有限公司 | Article recommendation method, device, medium and computer equipment |
CN111460130A (en) * | 2020-03-27 | 2020-07-28 | 咪咕数字传媒有限公司 | Information recommendation method, device, equipment and readable storage medium |
CN111476642A (en) * | 2020-04-14 | 2020-07-31 | 广东技术师范大学 | Insurance recommendation method and system based on social platform user head portrait analysis |
CN111489196A (en) * | 2020-04-08 | 2020-08-04 | 支付宝(杭州)信息技术有限公司 | Prediction method and device based on deep learning network, electronic equipment and medium |
CN111506803A (en) * | 2020-03-17 | 2020-08-07 | 北京百度网讯科技有限公司 | Content recommendation method and device, electronic equipment and storage medium |
CN111553763A (en) * | 2020-04-26 | 2020-08-18 | 上海风秩科技有限公司 | Article recommendation method and device, electronic equipment and readable storage medium |
CN111639199A (en) * | 2020-06-04 | 2020-09-08 | 北京声智科技有限公司 | Multimedia file recommendation method, device, server and storage medium |
CN111666496A (en) * | 2020-06-09 | 2020-09-15 | 中森云链(成都)科技有限责任公司 | Group recommendation method based on comment text |
CN112035747A (en) * | 2020-09-03 | 2020-12-04 | 腾讯科技(深圳)有限公司 | Information recommendation method and device |
CN112070586A (en) * | 2020-09-09 | 2020-12-11 | 腾讯科技(深圳)有限公司 | Article recommendation method and device based on semantic recognition, computer equipment and medium |
CN112114968A (en) * | 2020-09-18 | 2020-12-22 | 北京达佳互联信息技术有限公司 | Recommendation method and device, electronic equipment and storage medium |
CN112182351A (en) * | 2020-09-28 | 2021-01-05 | 哈尔滨工业大学(深圳) | News recommendation method and device based on multi-feature fusion |
CN112328646A (en) * | 2021-01-04 | 2021-02-05 | 平安科技(深圳)有限公司 | Multitask course recommendation method and device, computer equipment and storage medium |
CN112418423A (en) * | 2020-11-24 | 2021-02-26 | 百度在线网络技术(北京)有限公司 | Method, apparatus, and medium for recommending objects to a user using a neural network |
CN112597392A (en) * | 2020-12-25 | 2021-04-02 | 厦门大学 | Recommendation system based on dynamic attention and hierarchical reinforcement learning |
CN112632390A (en) * | 2020-12-29 | 2021-04-09 | 北京鸿享技术服务有限公司 | Information recommendation method, device and equipment based on label and storage medium |
WO2021077989A1 (en) * | 2019-10-25 | 2021-04-29 | 腾讯科技(深圳)有限公司 | Method and device for making recommendation, computer device, and storage medium |
CN112800097A (en) * | 2021-01-15 | 2021-05-14 | 稿定(厦门)科技有限公司 | Special topic recommendation method and device based on deep interest network |
CN112804080A (en) * | 2020-12-24 | 2021-05-14 | 中国科学院信息工程研究所 | Intelligent recommendation method for access control initialization |
CN112860999A (en) * | 2021-02-10 | 2021-05-28 | 脸萌有限公司 | Information recommendation method, device, equipment and storage medium |
CN112905879A (en) * | 2021-01-29 | 2021-06-04 | 北京达佳互联信息技术有限公司 | Recommendation method, recommendation device, server and storage medium |
CN113095433A (en) * | 2021-04-27 | 2021-07-09 | 北京石油化工学院 | Method for training intrusion detection network structure model |
CN113139834A (en) * | 2021-04-29 | 2021-07-20 | 北京沃东天骏信息技术有限公司 | Information processing method, device, electronic equipment and storage medium |
CN113157898A (en) * | 2021-05-26 | 2021-07-23 | 中国平安人寿保险股份有限公司 | Method and device for recommending candidate questions, computer equipment and storage medium |
WO2021159776A1 (en) * | 2020-02-13 | 2021-08-19 | 腾讯科技(深圳)有限公司 | Artificial intelligence-based recommendation method and apparatus, electronic device, and storage medium |
CN113379482A (en) * | 2021-05-28 | 2021-09-10 | 车智互联(北京)科技有限公司 | Item recommendation method, computing device and storage medium |
CN113393281A (en) * | 2020-03-11 | 2021-09-14 | 北京沃东天骏信息技术有限公司 | Method and device for processing request |
WO2021196470A1 (en) * | 2020-03-31 | 2021-10-07 | 百度在线网络技术(北京)有限公司 | Information pushing method and apparatus, device, and storage medium |
CN113518201A (en) * | 2020-07-14 | 2021-10-19 | 阿里巴巴集团控股有限公司 | Video processing method, device and equipment |
CN113705782A (en) * | 2021-08-18 | 2021-11-26 | 上海明略人工智能(集团)有限公司 | Model training method and device for media data recommendation |
CN113761392A (en) * | 2021-09-14 | 2021-12-07 | 上海任意门科技有限公司 | Content recall method, computing device, and computer-readable storage medium |
CN113836869A (en) * | 2021-09-22 | 2021-12-24 | 中国农业银行股份有限公司 | Method and device for carrying out unified code conversion on mixed multi-code character text |
CN113836439A (en) * | 2021-09-14 | 2021-12-24 | 上海任意门科技有限公司 | User matching method, computing device, and computer-readable storage medium |
US11669530B2 (en) | 2020-03-31 | 2023-06-06 | Baidu Online Network Technology (Beijing) Co., Ltd. | Information push method and apparatus, device, and storage medium |
TWI817921B (en) * | 2023-05-31 | 2023-10-01 | 明合智聯股份有限公司 | Model modeling instruction generation method and system |
CN111242318B (en) * | 2020-01-13 | 2024-04-26 | 拉扎斯网络科技(上海)有限公司 | Service model training method and device based on heterogeneous feature library |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103186527A (en) * | 2011-12-27 | 2013-07-03 | 北京百度网讯科技有限公司 | System for building music classification model, system for recommending music and corresponding method |
CN106959966A (en) * | 2016-01-12 | 2017-07-18 | 腾讯科技(深圳)有限公司 | A kind of information recommendation method and system |
CN109359244A (en) * | 2018-10-30 | 2019-02-19 | 中国科学院计算技术研究所 | A kind of recommendation method for personalized information and device |
-
2019
- 2019-04-23 CN CN201910326702.1A patent/CN110162700A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103186527A (en) * | 2011-12-27 | 2013-07-03 | 北京百度网讯科技有限公司 | System for building music classification model, system for recommending music and corresponding method |
CN106959966A (en) * | 2016-01-12 | 2017-07-18 | 腾讯科技(深圳)有限公司 | A kind of information recommendation method and system |
CN109359244A (en) * | 2018-10-30 | 2019-02-19 | 中国科学院计算技术研究所 | A kind of recommendation method for personalized information and device |
Cited By (67)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110781321A (en) * | 2019-08-28 | 2020-02-11 | 腾讯科技(深圳)有限公司 | Multimedia content recommendation method and device |
CN110781321B (en) * | 2019-08-28 | 2023-06-20 | 腾讯科技(深圳)有限公司 | Multimedia content recommendation method and device |
CN110704739A (en) * | 2019-09-30 | 2020-01-17 | 汉海信息技术(上海)有限公司 | Resource recommendation method and device and computer storage medium |
CN110766456A (en) * | 2019-10-16 | 2020-02-07 | 无线生活(杭州)信息科技有限公司 | Commodity recommendation method and device |
WO2021077989A1 (en) * | 2019-10-25 | 2021-04-29 | 腾讯科技(深圳)有限公司 | Method and device for making recommendation, computer device, and storage medium |
CN110928990A (en) * | 2019-10-31 | 2020-03-27 | 南方电网调峰调频发电有限公司 | Method special for recommending standing book data of power equipment based on user portrait |
CN110928990B (en) * | 2019-10-31 | 2023-08-22 | 南方电网调峰调频发电有限公司 | Method for recommending special power equipment ledger data based on user portrait |
CN111008321A (en) * | 2019-11-18 | 2020-04-14 | 广东技术师范大学 | Recommendation method and device based on logistic regression, computing equipment and readable storage medium |
CN111008321B (en) * | 2019-11-18 | 2023-08-29 | 广东技术师范大学 | Logistic regression recommendation-based method, device, computing equipment and readable storage medium |
CN111159564A (en) * | 2019-12-31 | 2020-05-15 | 联想(北京)有限公司 | Information recommendation method and device, storage medium and computer equipment |
CN111242318B (en) * | 2020-01-13 | 2024-04-26 | 拉扎斯网络科技(上海)有限公司 | Service model training method and device based on heterogeneous feature library |
CN111242318A (en) * | 2020-01-13 | 2020-06-05 | 拉扎斯网络科技(上海)有限公司 | Business model training method and device based on heterogeneous feature library |
CN111259222B (en) * | 2020-01-22 | 2023-08-22 | 北京百度网讯科技有限公司 | Article recommendation method, system, electronic equipment and storage medium |
CN111259222A (en) * | 2020-01-22 | 2020-06-09 | 北京百度网讯科技有限公司 | Article recommendation method, system, electronic device and storage medium |
WO2021159776A1 (en) * | 2020-02-13 | 2021-08-19 | 腾讯科技(深圳)有限公司 | Artificial intelligence-based recommendation method and apparatus, electronic device, and storage medium |
CN111369315A (en) * | 2020-02-27 | 2020-07-03 | 拉扎斯网络科技(上海)有限公司 | Resource object recommendation method and device, and data prediction model training method and device |
CN111368034A (en) * | 2020-03-02 | 2020-07-03 | 复旦大学 | Bidirectional semantic feature matching method and supply content recommendation device |
CN113393281A (en) * | 2020-03-11 | 2021-09-14 | 北京沃东天骏信息技术有限公司 | Method and device for processing request |
CN111506803B (en) * | 2020-03-17 | 2023-10-31 | 阿波罗智联(北京)科技有限公司 | Content recommendation method and device, electronic equipment and storage medium |
CN111506803A (en) * | 2020-03-17 | 2020-08-07 | 北京百度网讯科技有限公司 | Content recommendation method and device, electronic equipment and storage medium |
CN111400613A (en) * | 2020-03-17 | 2020-07-10 | 苏宁金融科技(南京)有限公司 | Article recommendation method, device, medium and computer equipment |
CN111460130A (en) * | 2020-03-27 | 2020-07-28 | 咪咕数字传媒有限公司 | Information recommendation method, device, equipment and readable storage medium |
US11669530B2 (en) | 2020-03-31 | 2023-06-06 | Baidu Online Network Technology (Beijing) Co., Ltd. | Information push method and apparatus, device, and storage medium |
WO2021196470A1 (en) * | 2020-03-31 | 2021-10-07 | 百度在线网络技术(北京)有限公司 | Information pushing method and apparatus, device, and storage medium |
CN111489196A (en) * | 2020-04-08 | 2020-08-04 | 支付宝(杭州)信息技术有限公司 | Prediction method and device based on deep learning network, electronic equipment and medium |
CN111476642A (en) * | 2020-04-14 | 2020-07-31 | 广东技术师范大学 | Insurance recommendation method and system based on social platform user head portrait analysis |
CN111553763A (en) * | 2020-04-26 | 2020-08-18 | 上海风秩科技有限公司 | Article recommendation method and device, electronic equipment and readable storage medium |
CN111553763B (en) * | 2020-04-26 | 2023-03-24 | 上海风秩科技有限公司 | Article recommendation method and device, electronic equipment and readable storage medium |
CN111639199A (en) * | 2020-06-04 | 2020-09-08 | 北京声智科技有限公司 | Multimedia file recommendation method, device, server and storage medium |
CN111666496B (en) * | 2020-06-09 | 2022-12-16 | 中森云链(成都)科技有限责任公司 | Group recommendation method based on comment text |
CN111666496A (en) * | 2020-06-09 | 2020-09-15 | 中森云链(成都)科技有限责任公司 | Group recommendation method based on comment text |
CN113518201A (en) * | 2020-07-14 | 2021-10-19 | 阿里巴巴集团控股有限公司 | Video processing method, device and equipment |
CN113518201B (en) * | 2020-07-14 | 2022-08-12 | 阿里巴巴集团控股有限公司 | Video processing method, device and equipment |
CN112035747A (en) * | 2020-09-03 | 2020-12-04 | 腾讯科技(深圳)有限公司 | Information recommendation method and device |
CN112035747B (en) * | 2020-09-03 | 2023-09-29 | 腾讯科技(深圳)有限公司 | Information recommendation method and device |
CN112070586B (en) * | 2020-09-09 | 2023-11-28 | 腾讯科技(深圳)有限公司 | Item recommendation method and device based on semantic recognition, computer equipment and medium |
CN112070586A (en) * | 2020-09-09 | 2020-12-11 | 腾讯科技(深圳)有限公司 | Article recommendation method and device based on semantic recognition, computer equipment and medium |
CN112114968A (en) * | 2020-09-18 | 2020-12-22 | 北京达佳互联信息技术有限公司 | Recommendation method and device, electronic equipment and storage medium |
CN112182351A (en) * | 2020-09-28 | 2021-01-05 | 哈尔滨工业大学(深圳) | News recommendation method and device based on multi-feature fusion |
CN112418423B (en) * | 2020-11-24 | 2023-08-15 | 百度在线网络技术(北京)有限公司 | Method, apparatus and medium for recommending objects to user using neural network |
CN112418423A (en) * | 2020-11-24 | 2021-02-26 | 百度在线网络技术(北京)有限公司 | Method, apparatus, and medium for recommending objects to a user using a neural network |
CN112804080A (en) * | 2020-12-24 | 2021-05-14 | 中国科学院信息工程研究所 | Intelligent recommendation method for access control initialization |
CN112804080B (en) * | 2020-12-24 | 2022-09-30 | 中国科学院信息工程研究所 | Intelligent recommendation method for access control initialization |
CN112597392A (en) * | 2020-12-25 | 2021-04-02 | 厦门大学 | Recommendation system based on dynamic attention and hierarchical reinforcement learning |
CN112632390A (en) * | 2020-12-29 | 2021-04-09 | 北京鸿享技术服务有限公司 | Information recommendation method, device and equipment based on label and storage medium |
CN112328646A (en) * | 2021-01-04 | 2021-02-05 | 平安科技(深圳)有限公司 | Multitask course recommendation method and device, computer equipment and storage medium |
CN112800097A (en) * | 2021-01-15 | 2021-05-14 | 稿定(厦门)科技有限公司 | Special topic recommendation method and device based on deep interest network |
CN113688167A (en) * | 2021-01-15 | 2021-11-23 | 稿定(厦门)科技有限公司 | Deep interest capture model construction method and device based on deep interest network |
CN112905879B (en) * | 2021-01-29 | 2024-03-19 | 北京达佳互联信息技术有限公司 | Recommendation method, recommendation device, server and storage medium |
CN112905879A (en) * | 2021-01-29 | 2021-06-04 | 北京达佳互联信息技术有限公司 | Recommendation method, recommendation device, server and storage medium |
CN112860999B (en) * | 2021-02-10 | 2023-04-18 | 脸萌有限公司 | Information recommendation method, device, equipment and storage medium |
CN112860999A (en) * | 2021-02-10 | 2021-05-28 | 脸萌有限公司 | Information recommendation method, device, equipment and storage medium |
CN113095433B (en) * | 2021-04-27 | 2023-06-23 | 北京石油化工学院 | Training method for intrusion detection network structure model |
CN113095433A (en) * | 2021-04-27 | 2021-07-09 | 北京石油化工学院 | Method for training intrusion detection network structure model |
CN113139834A (en) * | 2021-04-29 | 2021-07-20 | 北京沃东天骏信息技术有限公司 | Information processing method, device, electronic equipment and storage medium |
CN113157898A (en) * | 2021-05-26 | 2021-07-23 | 中国平安人寿保险股份有限公司 | Method and device for recommending candidate questions, computer equipment and storage medium |
CN113157898B (en) * | 2021-05-26 | 2022-10-14 | 中国平安人寿保险股份有限公司 | Method and device for recommending candidate questions, computer equipment and storage medium |
CN113379482A (en) * | 2021-05-28 | 2021-09-10 | 车智互联(北京)科技有限公司 | Item recommendation method, computing device and storage medium |
CN113379482B (en) * | 2021-05-28 | 2023-12-01 | 车智互联(北京)科技有限公司 | Article recommendation method, computing device and storage medium |
CN113705782A (en) * | 2021-08-18 | 2021-11-26 | 上海明略人工智能(集团)有限公司 | Model training method and device for media data recommendation |
CN113836439B (en) * | 2021-09-14 | 2024-01-30 | 上海任意门科技有限公司 | User matching method, computing device, and computer-readable storage medium |
CN113836439A (en) * | 2021-09-14 | 2021-12-24 | 上海任意门科技有限公司 | User matching method, computing device, and computer-readable storage medium |
CN113761392B (en) * | 2021-09-14 | 2022-04-12 | 上海任意门科技有限公司 | Content recall method, computing device, and computer-readable storage medium |
CN113761392A (en) * | 2021-09-14 | 2021-12-07 | 上海任意门科技有限公司 | Content recall method, computing device, and computer-readable storage medium |
CN113836869B (en) * | 2021-09-22 | 2023-12-08 | 中国农业银行股份有限公司 | Method and device for carrying out unified code conversion on hybrid multi-code character text |
CN113836869A (en) * | 2021-09-22 | 2021-12-24 | 中国农业银行股份有限公司 | Method and device for carrying out unified code conversion on mixed multi-code character text |
TWI817921B (en) * | 2023-05-31 | 2023-10-01 | 明合智聯股份有限公司 | Model modeling instruction generation method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110162700A (en) | The training method of information recommendation and model, device, equipment and storage medium | |
CN111177575B (en) | Content recommendation method and device, electronic equipment and storage medium | |
CN111681059B (en) | Training method and device of behavior prediction model | |
CN110909176B (en) | Data recommendation method and device, computer equipment and storage medium | |
CN109492157B (en) | News recommendation method and theme characterization method based on RNN and attention mechanism | |
Pan et al. | Study on convolutional neural network and its application in data mining and sales forecasting for E-commerce | |
CN110162703A (en) | Content recommendation method, training method, device, equipment and storage medium | |
CN110532479A (en) | A kind of information recommendation method, device and equipment | |
CN110008399A (en) | A kind of training method and device, a kind of recommended method and device of recommended models | |
CN102902691B (en) | Recommend method and system | |
CN111915395A (en) | Travel bag recommendation method based on multi-view attention mechanism | |
CN111898031B (en) | Method and device for obtaining user portrait | |
CN111784455A (en) | Article recommendation method and recommendation equipment | |
CN109509033A (en) | A kind of user buying behavior big data prediction technique under consumer finance scene | |
CN107077486A (en) | Affective Evaluation system and method | |
CN111159341B (en) | Information recommendation method and device based on user investment and financial management preference | |
CN110413894A (en) | The training method of commending contents model, the method for commending contents and relevant apparatus | |
CN109993583A (en) | Information-pushing method and device, storage medium and electronic device | |
CN110727864A (en) | User portrait method based on mobile phone App installation list | |
CN115860870A (en) | Commodity recommendation method, system and device and readable medium | |
CN115545832A (en) | Commodity search recommendation method and device, equipment and medium thereof | |
CN116823410A (en) | Data processing method, object processing method, recommending method and computing device | |
Velàsquez et al. | Building a knowledge base for implementing a web-based computerized recommendation system | |
CN111681107A (en) | Real-time personalized financial product recommendation algorithm based on Embedding | |
CN115525819A (en) | Cross-domain recommendation method for information cocoon room |
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 |