CN108989397A - Data recommendation method, device and storage medium - Google Patents
Data recommendation method, device and storage medium Download PDFInfo
- Publication number
- CN108989397A CN108989397A CN201810670441.0A CN201810670441A CN108989397A CN 108989397 A CN108989397 A CN 108989397A CN 201810670441 A CN201810670441 A CN 201810670441A CN 108989397 A CN108989397 A CN 108989397A
- Authority
- CN
- China
- Prior art keywords
- model
- data
- deepfm
- request
- default
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/50—Network services
- H04L67/55—Push-based network services
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/50—Network services
- H04L67/56—Provisioning of proxy services
- H04L67/565—Conversion or adaptation of application format or content
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Signal Processing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of data recommendation method, device and storage mediums, this method comprises: obtaining the data recommendation request of user, and are requested according to data recommendation, obtain the corresponding initial recommendation data of user and the corresponding characteristic of user;Data recommendation request is converted into the request of preset format, and according to the request of preset format, default DeepFM model is obtained from model server, wherein the request of preset format meets the identification requirement of model server;According to default DeepFM model and characteristic, more new data is generated;According to more new data and initial recommendation data, target recommending data is generated;By target recommending data, it is pushed to user.The program is based on default DeepFM model, generates more new data, and be updated to initial recommendation data according to more new data, generates target recommending data, improves the efficiency of data recommendation.
Description
Technical field
The present invention relates to field of communication technology more particularly to a kind of data recommendation methods, device and storage medium.
Background technique
With popularizing for the mobile terminals such as smart phone, tablet computer, user can be obtained more and more by terminal
Information.In order to make user's quick obtaining to desired information, information recommendation can be carried out according to the hobby of user.
Logistic regression (Logistic Regression, abbreviation LR) algorithm is generally used in existing recommender system, LR is calculated
Method has complexity lower, the advantages of should be readily appreciated that and realize, but also has limitation.LR algorithm can only be automatically to low order spy
Sign is described, and during combining high-order feature, needs manually to carry out feature combination.And the work in feature anabolic process
Measure it is larger, cause recommend efficiency be lower.
Summary of the invention
The embodiment of the present invention provides a kind of data recommendation method, device and storage medium, and the effect of data recommendation can be improved
Rate.
The embodiment of the present invention provides a kind of data recommendation method, comprising: obtains the data recommendation request of user, and according to institute
Data recommendation request is stated, the corresponding initial recommendation data of the user and the corresponding characteristic of the user are obtained;By institute
The request that data recommendation request is converted into preset format is stated, and according to the request of the preset format, is obtained from model server
Default DeepFM model is taken, wherein the request of the preset format meets the identification requirement of the model server;According to described
Default DeepFM model and the characteristic generate more new data;According to the more new data and the initial recommendation
Data generate target recommending data;By the target recommending data, it is pushed to the user.
The embodiment of the present invention also provides a kind of data recommendation device, comprising: first obtains module, for obtaining the number of user
It is requested according to recommendation request, and according to the data recommendation, obtains the corresponding initial recommendation data of the user and the user
Corresponding characteristic;Conversion module, for data recommendation request to be converted into the request of preset format, and according to described
The request of preset format obtains default DeepFM model, wherein the request of the preset format meets institute from model server
State the identification requirement of model server;First generation module, for according to the default DeepFM model and the characteristic
According to generation more new data;Second generation module, for generating mesh according to the more new data and the initial recommendation data
Mark recommending data;Pushing module, for being pushed to the user for the target recommending data.
The embodiment of the present invention also provides a kind of storage medium, is stored with processor-executable instruction, and the processor is logical
It crosses and described instruction offer such as above-mentioned data recommendation method is provided.
Data recommendation method, device and the storage medium of the embodiment of the present invention, first pass through according to preset DeepFM model with
And characteristic, more new data is generated, further according to more new data and initial recommendation data, generates target recommending data, finally
Target recommending data is pushed to user, improves the efficiency of data recommendation.
Detailed description of the invention
With reference to the accompanying drawing, by the way that detailed description of specific embodiments of the present invention, technical solution of the present invention will be made
And other beneficial effects are apparent.
Fig. 1 is the schematic diagram of a scenario of data recommendation method provided in an embodiment of the present invention.
Fig. 2 is the flow diagram of data recommendation method provided in an embodiment of the present invention.
Fig. 3 is another flow diagram of data recommendation method provided in an embodiment of the present invention.
Fig. 4 is the structural schematic diagram of data recommendation device provided in an embodiment of the present invention.
Fig. 5 is the structural schematic diagram of generation module provided in an embodiment of the present invention.
Fig. 6 is the structural schematic diagram of conversion module provided in an embodiment of the present invention.
Fig. 7 is the structural schematic diagram of the first generation module provided in an embodiment of the present invention.
Fig. 8 is the structural schematic diagram of electronic equipment provided in an embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those skilled in the art's every other implementation obtained without creative efforts
Example, shall fall within the protection scope of the present invention.
Referring to Fig. 1, which is the schematic diagram of a scenario of data recommendation method provided in an embodiment of the present invention, in the scene, number
It can be used as entity according to recommendation apparatus to realize, also can integrate and realized in the electronic equipments such as terminal or server, the electronics
Equipment may include smart phone, tablet computer and personal computer etc..
As shown in Figure 1, may include terminal a, server b, server c, server d and server e in the scene.Its
In, server c is used for storing initial for storing characteristic, server e for storing default DeepFM model, server d
Recommending data.Specifically, user A generates data recommendation request by operating terminal a, which is requested to send by terminal a
Give server b.After server b gets data recommendation request, characteristic relevant to user A successively is obtained from server d
According to the acquisition initial recommendation data relevant to user A from server e.Further, server b also asks the data recommendation
The request for being converted into preset format is asked, to meet the identification requirement of server c.The request of the preset format is sent to service again
Device c obtains default DeepFM model from server c.Then, server b presets DeepFM model and user's A phase according to this
The characteristic of pass generates more new data.Further according to more new data initial recommendation data relevant with the user A, mesh is generated
Mark recommending data.Finally, the target recommending data is sent to terminal a again, user A is showed by terminal a.
The embodiment of the present invention provides a kind of data recommendation method, device and storage medium, will carry out respectively specifically below
It is bright.
In embodiments of the present invention, it will be described from the angle of data recommendation device, which specifically may be used
To integrate in the electronic device.
A kind of data recommendation method including obtaining the data recommendation request of user, and requests according to data recommendation, obtains and use
The corresponding initial recommendation data in family and the corresponding characteristic of user;Data recommendation request is converted into asking for preset format
It asks, and according to the request of preset format, default DeepFM model is obtained from model server, wherein the request of preset format is full
The identification requirement of sufficient model server;According to default DeepFM model and characteristic, more new data is generated;According to update number
According to and initial recommendation data, generate target recommending data;By target recommending data, it is pushed to user.
Referring to figure 2., Fig. 2 is the flow chart of data recommendation method provided in an embodiment of the present invention, and this method may include:
Step S101, obtains the data recommendation request of user, and is requested according to data recommendation, and it is corresponding initial to obtain user
Recommending data and the corresponding characteristic of user.
When user carries out relevant operation to terminal, data recommendation request can be generated.Wherein, which includes updating behaviour
Work, inquiry operation etc..For example, when clicking " carrying out a head again " button on mobile phone interface, will be touched when user listens song using mobile phone
Data recommendation occurs into request, then terminal obtains data recommendation request, is requested according to the data recommendation, obtains the user couple
The initial recommendation data and the corresponding characteristic of the user answered are suitable to recommend to user to analyze the hobby of user
Song track.
Initial recommendation data refer to the recommending data prestored, such as list of songs, list of videos etc..A storage can be set
Area comes storing initial recommending data, user identifier and incidence relation between the two.When the data recommendation for getting user is asked
After asking, can be searched from the memory block with the matched user identifier of the user, then obtain that the user identifier is corresponding initially to be pushed away
Recommend data.
Characteristic can be extracted from user information by collecting user information.Specifically, can be from user information
The information such as attribute information, operation information and the corresponding contextual information of the operation information of user are chosen, to extract characteristic
According to.
Wherein, the attribute information of user refers to biological nature of the people under natural law effect, and in social environment and
Social property under conditioning, what is embodied is the intrinsic property and feature of people.Specifically, the ascribed characteristics of population includes the property of user
Not, the information such as age and region.Operation information refers to the information generated during user and terminal interaction, specifically includes: behaviour
Make type, operating time information and operation object etc..For example, operation object can be right for audio, video and text etc.
The action type of audio includes deleting audio, collection audio and downloading audio etc., and the operating time information to audio includes deleting
Except the time of audio, the duration of collection audio and the time for downloading audio etc..The corresponding contextual information of operation information refers to
Information associated with operation information, such as operating system, entrance when listening to audio of terminal etc..For example, listening to a certain song
Qu Shi can input song title entrance in search box, can also enter from the song of collection.
Data recommendation request is converted into the request of preset format by step S102, and according to the request of preset format, from mould
Default DeepFM model is obtained on type server, wherein the request of preset format meets the identification requirement of model server.
Default DeepFM model is to be generated using data set to the first DeepFM model training.Wherein the first DeepFM mould
There is parameter, during training, the parameter value of the available parameter in type.The parameter value is substituted into the first DeepFM
Default DeepFM model can be obtained in model.
It should be noted that the first DeepFM model is by FM (Factorization Machine, Factorization machine) model
It is generated with DNN (Deep Neural Networks, deep neural network) models coupling.Wherein, FM model is a kind of based on square
The machine learning algorithm that battle array is decomposed, carrys out transformation factor decomposition model using feature vector, passes through the inner product between low order relationship
The interactive relation being expressed as between the factor.FM model has the advantages that automatic assemblage characteristic, so as to largely reduce feature work
The work of journey.DNN model can combine the combination of high-order feature and the combination of low order feature, in the present embodiment, first
DeepFM model can automatically generate low order feature and high-order feature, mention in conjunction with the advantage of both DNN model and FM model
Treatment effeciency in terms of high Feature Engineering.
Above-mentioned default DeepFM model is arranged on model server, such as TFServing server.Specifically, can be with
Default DeepFM model is configured under the specified catalogue of TFServing server, so that the load of TFServing server should
Default DeepFM model.When requesting to service to model servers such as TFServing servers, due to different model servers
Its agreement supported is different, it is possible that the case where model server cannot identify data recommendation request.Specifically, assuming mould
Type server only identifies the request of gRPC format, then when data recommendation request is JSON (JavaScript Object
Notation) when the request of format, model server can be requested because that cannot accurately identify the data recommendation of the JSON format, without
Default DeepFM model information can be provided in time.To sum up, it needs to format data recommendation request, to meet model clothes
The identification requirement of business device obtains default DeepFM model further according to the request of preset format from model server.
Step S103 generates more new data according to default DeepFM model and characteristic.
Specifically, can be first using the FM model in default DeepFM model, by characteristic vectorization, then by vector
The characteristic of change inputs DNN model as input, generates the first output result.FM model is inputted with regard to characteristic again, is generated
Second output result.Finally, the first output result and second is exported the input as a result, as sigmoid function, to obtain
More new data.
Step S104 generates target recommending data according to more new data and initial recommendation data.
In some embodiments, the more new data can be used, initial recommendation data are updated, to generate target
Recommending data.It is assumed that initial recommendation data include a song recommendations list, more new data is also a song recommendations list.Then may be used
To search the song being overlapped in initial recommendation data and more new data, and target recommending data is set by the song of coincidence.
It in some embodiments,, may after the data recommendation request for getting user due to being limited by computing capability
It is unable to the request of the real-time response data recommendation, it is possible to because calculating overlong time can cause that more new data cannot be generated in time
The case where.In the case where the more new data of above-mentioned generation is empty, directly number can be recommended using initial recommendation data as target
According in so avoiding users to wait data-pushing for a long time excessively, also reduces the pressure of calculating.
Target recommending data is pushed to user by step S105.
Specifically, can be by showing target recommending data on terminal interface, realization pushes target recommending data
To the purpose of user.
It can be seen from the above, the data recommendation method of the embodiment of the present invention, first passes through according to default DeepFM model and spy
Data are levied, more new data is generated, further according to more new data and initial recommendation data, target recommending data are generated, finally by mesh
Mark recommending data is pushed to user, improves the efficiency of data recommendation.
According to the data recommendation method that above-described embodiment describes, citing is described further below.Implement in the present invention
In example, it will be described from the angle of data recommendation device, which specifically can integrate in the electronic device.
Referring to figure 3., Fig. 3 is another flow chart of data recommendation method provided in an embodiment of the present invention, and this method can be with
Include:
Step S201 obtains the history feature data of user in real time.
With the development of internet, it is got worse the phenomenon that information overload.And the recommended method based on machine learning, it can be with
The selection such as suitable commodity, music and film is provided for user, therefore is widely used.The wherein key of machine learning
It is the selection of model and characteristic.In the present embodiment, it can extract and go through from user information by collecting user information
History characteristic.Specifically, attribute information, historical operation information and the history behaviour of user can be chosen from user information
Make the information such as the corresponding contextual information of information, to extract the history feature data for model training.
Wherein, the attribute information of user refers to biological nature of the people under natural law effect, and in social environment and
Social property under conditioning, what is embodied is the intrinsic property and feature of people.Specifically, the ascribed characteristics of population includes the property of user
Not, the information such as age and region.Historical operation information refers to the information generated during user and terminal interaction, specific to wrap
It includes: action type, operating time information and operation object etc..For example, operation object can be audio, video and text
Deng including deleting audio, collection audio and downloading audio etc. to the action type of audio, to the operating time packet of audio
It includes and deletes the time of audio, the duration for collecting audio and the time for downloading audio etc..The corresponding context of historical operation information
Information refers to information associated with historical operation information, such as operating system, entrance when listening to audio of terminal etc..Example
Such as, when listening to a certain song, song title entrance can be inputted in search box, can also be entered from the song of collection.
It should be noted that since user and terminal are in continuous interactive process, correspondingly, user information also constantly exists
It updates, therefore, it is necessary to obtain history feature data from user information in real time.In some embodiments, it can use and deposit offline
The mode of storage stores history feature data.In this way can in off-line case, usage history characteristic is to the first DeepFM mould
Type is trained, and to reduce network overhead, is detailed in following steps S202.
Step S202, usage history characteristic are trained the first DeepFM model, generate training result.
First DeepFM model is by FM (Factorization Machine, Factorization machine) model and DNN (Deep
Neural Networks, deep neural network) models coupling generation.Wherein, FM model is a kind of machine based on matrix decomposition
Learning algorithm carrys out transformation factor decomposition model using feature vector, by the inner product between low order relationship be expressed as the factor it
Between interactive relation.FM model has the advantages that automatic assemblage characteristic, can largely reduce the work of Feature Engineering.DNN model
The combination of high-order feature and the combination of low order feature can be combined.In the present embodiment, using the first DeepFM model, Ke Yijie
The advantage for closing both DNN model and FM model, can automatically generate low order feature and high-order feature, in terms of improving Feature Engineering
Treatment effeciency.
There is parameter, by using above-mentioned history feature data, on TensorFlow cluster in first DeepFM model
The first DeepFM model of training obtains the corresponding parameter value of the first DeepFM Model Parameter to get training result is arrived.
According to step S201 it is found that the history feature data of user are constantly changing, therefore obtained in real time by step S201
The history feature data got, the first DeepFM model of training, can also obtain multiple training results.
Step S203 generates multiple 2nd DeepFM models according to multiple training results and the first DeepFM model.
Specifically, can substitute into above-mentioned first DeepFM model using the training result as parameter value, the can be obtained
Two DeepFM models.It should be noted that due to multiple training results, correspondingly, can also generate multiple 2nd DeepFM
Model.
Step S204 obtains the version information of multiple 2nd DeepFM models.
Step S203 is accepted, training result is different, and it is also different to correspond to the 2nd DeepFM model generated.Multiple second
In DeepFM model, according to the 2nd DeepFM model being newly generated, the target recommending data of generation, accuracy is higher.
To sum up, in order to improve the accuracy of data recommendation, need to obtain the version information of multiple 2nd DeepFM model,
From multiple 2nd DeepFM model, the 2nd DeepFM model of latest edition is picked out.Preparatory, it can be according to second
The temporal information that DeepFM model generates carries out edition protection to multiple 2nd DeepFM model.Wherein, second is default
The time that DeepFM model generates is shorter, and version is higher, and the time of generation is longer, and version is lower.
Specifically, the temporal information includes timestamp, wherein timestamp be indicate a data some specific time it
Preceding already existing, the complete, data that can verify that, a usually character string, with the time at unique identification a moment.It is false
If some specific time is on January 1st, 1970, the generation time of the 2nd DeepFM model is 10:04 on April 10th, 2018:
35, corresponding timestamp is second to set generation time of DeepFM model to the number of seconds between the specific time --
1523325885.To sum up, can use 1523325885 as the 2nd DeepFM model version number, with unique identification this second
DeepFM model.It to sum up, can be by the version number of the multiple 2nd DeepFM models of acquisition, to obtain version information.
Step S205 from multiple 2nd DeepFM models, selects default DeepFM model according to version information, and
Default DeepFM model is arranged on TFServing server.
Specifically, accepting step S204, version number maximum second can be chosen from multiple 2nd DeepDM model
DeepFM model presets DeepFM model as this.Then by default DeepFM model, it is specified to be configured to TFServing server
Catalogue under, so that TFServing server loads the default DeepFM model.
It should be noted that since TFServing server service is the service of CPU intensive type, in order to realize
The optimization of resource allocation on TFServing server, can make TFServing server at regular intervals, remove specified mesh
Under record, inquiry is with the presence or absence of default DeepFM model, if so, then loading the default DeepFM model.
Step S206, obtains the data recommendation request of user, and is requested according to data recommendation, and it is corresponding initial to obtain user
Recommending data and the corresponding characteristic of user.
Wherein, initial recommendation data refer to the recommending data prestored, such as list of songs, list of videos etc..It can be set
Storing initial recommending data, user identifier and incidence relation between the two are come in one memory block.When the data for getting user
After recommendation request, can be searched from the memory block with the matched user identifier of the user, then to obtain the user identifier corresponding
Initial recommendation data.
In some embodiments, the corresponding characteristic of the user is using storage mode on line.Specifically, can use
The databases such as myslq or oracle store this feature data.Preferably, can also using CMEM (Cloud Memcache, it is interior
Deposit persistent storage) it is stored, CMEM stores characteristic using Key-Value form, can accelerate the reading of characteristic
It takes.
Step S207, is used as middleware services using Proxy service, and data recommendation request is converted into asking for gRPC format
It asks;By the request of gRPC format, from TFServing server, default DeepFM model is obtained.
After setting up default DeepFM model on TFServing server, if requesting to take to TFServing server
Business is needed using the gRPC protocol access Predict function in Proto3.Wherein, gRPC is a high-performance, general open source
RPC frame.Since the language that Proto3 agreement is supported is limited, in the actual operation process, it is possible that cannot use
The case where gRPC agreement.Therefore use Proxy service in the present embodiment and be used as middleware, by gRPC protocol conversion at general
Agreement.
To sum up, it is requested according to data recommendation, from TFServing server, obtains the step of presetting DeepFM model, tool
Body is as follows:
Assuming that data recommendation request is the request of JSON (JavaScript Object Notation) format, by the data
After recommendation request is sent to Proxy service in the form of SOKCET message, the data recommendation for parsing JSON format is asked in Proxy service
It asks, and is encapsulated into the request of gRPC format.Then, the request of gRPC format is sent to TFServing clothes by Proxy service
It is engaged in device, and receives the default DeepFM model relevant information of TFServing server return, then by the phase of default DeepFM model
It closes Information encapsulation to return at SOKCET message, it can get default DeepFM model.In some embodiments, it can incite somebody to action
Proxy service and TFServing service are arranged on same server, to reduce the time delay of data recommendation request.
It should be noted that due to TFServing server using multithreading run, receive data recommendation request and
Loading default two operations of DeepFM model can carry out simultaneously, will not mutually interrupt.
Step S208 generates more new data according to default DeepFM model and characteristic.
Due to presetting the combination that DeepFM model is FM model and DNN model, FM model, DNN model can be combined
And characteristic, generate more new data, the specific steps are as follows:
(A) FM model is used, by characteristic vectorization.
(B) by the characteristic of vectorization, DNN model is inputted, generates the first output result.
(C) according to the first output as a result, and FM model corresponding second exports as a result, generation more new data.
Specifically, after getting the first output result and the second output result, it can be in conjunction with sigmoid function, next life
At more new data.Assuming that the first output result is y1, the second output result is y2, then more new data Y=sigmoid (w1*y1+
W2*y2+bias), wherein w1, w2 and bias are parameter, and history feature data can be used in the value of above-mentioned parameter, to first
DeepFM model is trained generation, specifically as shown in step S202.
Step S209 generates target recommending data according to more new data and initial recommendation data.
In some embodiments, the more new data can be used, initial recommendation data are updated, to generate target
Recommending data.It is assumed that initial recommendation data include a song recommendations list, more new data is also a song recommendations list.Then may be used
To search the song being overlapped in initial recommendation data and more new data, and target recommending data is set by the song of coincidence.
It in some embodiments,, may after the data recommendation request for getting user due to being limited by computing capability
It is unable to the request of the real-time response data recommendation, it is possible to because calculating overlong time can cause that more new data cannot be generated in time
The case where.In the case where the more new data of above-mentioned generation is empty, directly number can be recommended using initial recommendation data as target
According in so avoiding users to wait data-pushing for a long time excessively, also reduces the pressure of calculating.
Target recommending data is pushed to user by step S210.
Specifically, can be by showing target recommending data on terminal interface, realization pushes target recommending data
To the purpose of user.
The data recommendation method of the embodiment of the present invention is first passed through according to DeepFM model and characteristic is preset, is generated
More new data generates target recommending data, finally pushes away target recommending data further according to more new data and initial recommendation data
User is given, the efficiency of data recommendation is improved.
The method according to described in above-described embodiment, the present embodiment will be retouched from the angle further progress of data recommendation device
It states, which can integrate in the electronic device
Referring to figure 4., Fig. 4 is the structure chart of data recommendation device provided in an embodiment of the present invention, which can wrap
Include the first acquisition module 301, conversion module 302, the first generation module 303, the second generation module 304 and pushing module 305.
(1) first obtains module 301
First obtains module 301, and the data recommendation for obtaining user is requested, and is requested according to data recommendation, obtains pre-
If DeepFM model, the corresponding initial recommendation data of user and the corresponding characteristic of user.
When user carries out relevant operation to terminal, data recommendation request can be generated.Wherein, which includes updating behaviour
Work, inquiry operation etc..For example, when clicking " carrying out a head again " button on mobile phone interface, triggering is generated when listening song using mobile phone
Then data recommendation request first obtains the acquisition data recommendation request of module 301, is requested, obtained pre- according to the data recommendation
If DeepFM model, the corresponding initial recommendation data of the user and the corresponding characteristic of the user, to analyze the happiness of user
It is good, to recommend suitable song track to user.
Wherein, initial recommendation data refer to the recommending data prestored, such as list of songs, list of videos etc..It can be set
Storing initial recommending data, user identifier and incidence relation between the two are come in one memory block.When the first acquisition module 301 obtains
After getting the data recommendation request of user, can be searched from the memory block and the matched user identifier of the user, then obtain this
The corresponding initial recommendation data of user identifier.
Characteristic can be extracted from user information by collecting user information.Specifically, first obtains module 301
The letter such as attribute information, operation information and corresponding contextual information of the operation information of user can be chosen from user information
Breath, to extract characteristic.
In some embodiments, the corresponding characteristic of the user is using storage mode on line.Specifically, can use
The databases such as myslq or oracle store this feature data.Preferably, can also using CMEM (Cloud Memcache, it is interior
Deposit persistent storage) it is stored, CMEM stores characteristic using Key-Value form, can accelerate the first acquisition module
301 read characteristic.
Default DeepFM model is to be generated using data set to the first DeepFM model training.Wherein the first DeepFM mould
There is parameter, during training, the parameter value of the available parameter in type.The parameter value is substituted into the first DeepFM
Default DeepFM model can be obtained in model.
It should be noted that the first DeepFM model is by FM (Factorization Machine, Factorization machine) model
It is generated with DNN (Deep Neural Networks, deep neural network) models coupling.Wherein, FM model is a kind of based on square
The machine learning algorithm that battle array is decomposed, carrys out transformation factor decomposition model using feature vector, passes through the inner product between low order relationship
The interactive relation being expressed as between the factor.FM model has the advantages that automatic assemblage characteristic, so as to largely reduce feature work
The work of journey.DNN model can combine the combination of high-order feature and the combination of low order feature, in the present embodiment, first
DeepFM model can automatically generate low order feature and high-order feature, mention in conjunction with the advantage of both DNN model and FM model
Treatment effeciency in terms of high Feature Engineering.
The generating process of default DeepFM model is described in detail below:
In some embodiments, as shown in figure 4, device 30 further include: second obtains module 306, training module 307 and life
At module 308.
Second obtains module 306 for obtaining the history feature data of user in real time.In some embodiments, it second obtains
Module 306 can extract history feature data by collecting user information from user information.Specifically, second obtains module
306 can from chosen in user information the attribute information of user, historical operation information and the historical operation information it is corresponding on
The information such as context information, to extract the history feature data for model training.
Wherein, the attribute information of user refers to biological nature of the people under natural law effect, and in social environment and
Social property under conditioning, what is embodied is the intrinsic property and feature of people.Specifically, the ascribed characteristics of population includes the property of user
Not, the information such as age and region.Historical operation information refers to the information generated during user and terminal interaction, specific to wrap
It includes: action type, operating time information and operation object etc..For example, operation object can be audio, video and text
Deng including deleting audio, collection audio and downloading audio etc. to the action type of audio, to the operating time packet of audio
It includes and deletes the time of audio, the duration for collecting audio and the time for downloading audio etc..The corresponding context of historical operation information
Information refers to information associated with historical operation information, such as operating system, entrance when listening to audio of terminal etc..Example
Such as, when listening to a certain song, song title entrance can be inputted in search box, can also be entered from the song of collection.
It should be noted that since user and terminal are in continuous interactive process, correspondingly, user information also constantly exists
It updates, obtains module 306 therefore, it is necessary to second and obtain history feature data from user information in real time.In some embodiments,
History feature data can be stored by the way of offline storage.Training module 307 can in off-line case, using going through in this way
History characteristic is trained the first DeepFM model, to reduce network overhead.
Training module 307 is described in detail below.Training module 307 is used for usage history characteristic, to the first DeepFM
Model is trained, and generates multiple training results.There is parameter, training module 307 is by using upper in first DeepFM model
History feature data are stated, the first DeepFM model of training, obtains joining in the first DeepFM model on TensorFlow cluster
Corresponding parameter value is counted to get training result is arrived.
Generation module 308 is used to generate default DeepFM model according to multiple training results and the first DeepFM model.
In some embodiments, as shown in figure 5, generation module 308 include: generate submodule 3081, the first acquisition submodule 3082 with
And choose submodule 3083.
Submodule 3081 is generated, for generating multiple second according to multiple training results and the first DeepFM model
DeepFM model.Specifically, above-mentioned first DeepFM can be substituted into using the training result as parameter value by generating submodule 3081
In model, the 2nd DeepFM model can be obtained.It should be noted that due to multiple training results, correspondingly, can also give birth to
At multiple 2nd DeepFM models.
First acquisition submodule 3082 is used to obtain the version information of multiple 2nd DeepFM models.Training result is different,
It is also different to generate corresponding the 2nd DeepFM model generated of submodule 3081.In multiple 2nd DeepFM models, according to nearest
The 2nd DeepFM model generated, the target recommending data of generation, accuracy are higher.
To sum up, in order to improve the accuracy of data recommendation, need to obtain the version information of multiple 2nd DeepFM model,
From multiple 2nd DeepFM model, the 2nd DeepFM model of latest edition is picked out.Preparatory, it can be according to second
The temporal information that DeepFM model generates carries out edition protection to multiple 2nd DeepFM model.Wherein, second is default
The time that DeepFM model generates is shorter, and version is higher, and the time of generation is longer, and version is lower.
Specifically, the temporal information includes timestamp, wherein timestamp be indicate a data some specific time it
Preceding already existing, the complete, data that can verify that, a usually character string, with the time at unique identification a moment.It is false
If some specific time is on January 1st, 1970, the generation time of the 2nd DeepFM model is 10:04 on April 10th, 2018:
35, corresponding timestamp is second to set generation time of DeepFM model to the number of seconds between the specific time --
1523325885.To sum up, can use 1523325885 as the 2nd DeepFM model version number, with unique identification this second
DeepFM model.To sum up, the first acquisition submodule 3082 can be come by the version number of the multiple 2nd DeepFM models of acquisition
To version information.
Submodule 3083 is chosen, for selecting default from multiple 2nd DeepFM models according to version information
DeepFM model.Specifically, choosing submodule 3083 it is maximum can to choose version number from multiple 2nd DeepDM model
2nd DeepFM model presets DeepFM model as this.
(2) conversion module 302
In some embodiments, as shown in fig. 6, conversion module 302 includes that transform subblock 3021 and second obtain submodule
Block 3022.Data recommendation request is converted by transform subblock 3021 for being used as middleware services using Proxy service
The request of gRPC format;Second acquisition submodule 3022, for passing through the request of gRPC format, from TFServing server,
Obtain default DeepFM model.
Specifically, assuming that data recommendation request is the request of JSON format, transform subblock 3021 asks the data recommendation
It asks after being sent to Proxy service in the form of SOKCET message, the data recommendation for parsing JSON format is requested in Proxy service, and
It is encapsulated into the request of gRPC format.Then, the request of gRPC format is sent to TFServing server by Proxy service,
And the default DeepFM model relevant information of TFServing server return is received, then the correlation of default DeepFM model is believed
Breath is packaged into the return of SOKCET message, i.e. the second acquisition submodule 3022 is available to default DeepFM model.In some realities
It applies in example, Proxy can be serviced and TFServing service is arranged on same server, to reduce data recommendation request
Time delay.
It should be noted that due to TFServing server using multithreading run, receive data recommendation request and
Loading default two operations of DeepFM model can carry out simultaneously, will not mutually interrupt.
It should be noted that since TFServing server service is the service of CPU intensive type, in order to realize
The optimization of resource allocation on TFServing server, can make TFServing server at regular intervals, remove specified mesh
Under record, inquiry is with the presence or absence of default DeepFM model, if so, then loading the default DeepFM model.
(3) first generation modules 303
First generation module 303, for generating more new data according to DeepFM model and characteristic is preset.It is default
DeepFM model includes FM model and DNN model, as shown in fig. 7, the first generation module 303 include: vectorization submodule 3021,
Input submodule 3022 and update submodule 3023.
Vectorization submodule 3021, for using FM model, by characteristic vectorization;Input submodule 3022, is used for
By the characteristic of vectorization, DNN model is inputted, generates the first output result;Submodule 3023 is updated, for defeated according to first
Out as a result, and FM model corresponding second exports as a result, generation more new data.
Specifically, updating submodule 3023 can combine after getting the first output result and the second output result
Sigmoid function, to generate update data.Assuming that the first output result is y1, the second output result is y2, then more new data Y
=sigmoid (w1*y1+w2*y2+bias), wherein w1, w2 and bias are parameter, and history can be used in the value of above-mentioned parameter
Characteristic is trained generation to the first DeepFM model.
(4) second generation modules 304
Second generation module 304, for generating target recommending data according to more new data and initial recommendation data.?
In some embodiments, the more new data is can be used in the second generation module 304, is updated to initial recommendation data, thus raw
At target recommending data.It is assumed that initial recommendation data include a song recommendations list, more new data is also song recommendations column
Table.Then the second generation module 304 can search the song being overlapped in initial recommendation data and more new data, and by the song of coincidence
It is set as target recommending data.
It in some embodiments,, may after the data recommendation request for getting user due to being limited by computing capability
It is unable to the request of the real-time response data recommendation, it is possible to because calculating overlong time can cause that more new data cannot be generated in time
The case where.In the case where the more new data of above-mentioned generation is empty, the second generation module 304 can be directly by initial recommendation data
As target recommending data, in so avoiding users to wait data-pushing for a long time excessively, also reduces the pressure of calculating.
(4) pushing module 305
Pushing module 305, for being pushed to user for target recommending data.Specifically, pushing module 305 can pass through
Target recommending data is shown on terminal interface, realizes the purpose that target recommending data is pushed to user.
The data recommendation device of the embodiment of the present invention is first passed through according to DeepFM model and characteristic is preset, is generated
More new data generates target recommending data, finally pushes away target recommending data further according to more new data and initial recommendation data
User is given, the efficiency of data recommendation is improved.
Correspondingly, the embodiment of the present invention also provides a kind of electronic equipment, as shown in figure 8, it illustrates the embodiment of the present invention
The structural schematic diagram of related electronic equipment, specifically:
The electronic equipment may include one or more than one processing core processor 401, one or more
The components such as memory 402, power supply 403 and the input unit 404 of computer readable storage medium.Those skilled in the art can manage
It solves, electronic devices structure shown in Fig. 8 does not constitute the restriction to electronic equipment, may include more more or fewer than illustrating
Component perhaps combines certain components or different component layouts.Wherein:
Processor 401 is the control centre of the electronic equipment, utilizes various interfaces and the entire electronic equipment of connection
Various pieces by running or execute the software program and/or module that are stored in memory 402, and are called and are stored in
Data in reservoir 402 execute the various functions and processing data of electronic equipment, to carry out integral monitoring to electronic equipment.
Optionally, processor 401 may include one or more processing cores;Preferably, processor 401 can integrate application processor and tune
Demodulation processor processed, wherein the main processing operation system of application processor, user interface and application program etc., modulatedemodulate is mediated
Reason device mainly handles wireless communication.It is understood that above-mentioned modem processor can not also be integrated into processor 401
In.
Memory 402 can be used for storing software program and module, and processor 401 is stored in memory 402 by operation
Software program and module, thereby executing various function application and data processing.Memory 402 can mainly include storage journey
Sequence area and storage data area, wherein storing program area can the (ratio of application program needed for storage program area, at least one function
Such as sound-playing function, image player function) etc.;Storage data area, which can be stored, uses created number according to electronic equipment
According to etc..In addition, memory 402 may include high-speed random access memory, it can also include nonvolatile memory, such as extremely
A few disk memory, flush memory device or other volatile solid-state parts.Correspondingly, memory 402 can also wrap
Memory Controller is included, to provide access of the processor 401 to memory 402.
Electronic equipment further includes the power supply 403 powered to all parts, it is preferred that power supply 403 can pass through power management
System and processor 401 are logically contiguous, to realize management charging, electric discharge and power managed etc. by power-supply management system
Function.Power supply 403 can also include one or more direct current or AC power source, recharging system, power failure monitor
The random components such as circuit, power adapter or inverter, power supply status indicator.
The electronic equipment may also include input unit 404, which can be used for receiving the number or character of input
Information, and generate keyboard related with user setting and function control, mouse, operating stick, optics or trackball signal
Input.
Although being not shown, electronic equipment can also be including display unit etc., and details are not described herein.Specifically in the present embodiment
In, the processor 401 in electronic equipment can be corresponding by the process of one or more application program according to following instruction
Executable file be loaded into memory 402, and the application program being stored in memory 402 is run by processor 401,
It is as follows to realize various functions:
The data recommendation request of user is obtained, and is requested according to data recommendation, the corresponding initial recommendation data of user are obtained,
And the corresponding characteristic of user;
Data recommendation request is converted into the request of preset format, and according to the request of preset format, from model server
Upper to obtain default DeepFM model, wherein the request of preset format meets the identification requirement of model server;
According to default DeepFM model and characteristic, more new data is generated;
According to more new data and initial recommendation data, target recommending data is generated;
By target recommending data, it is pushed to user.
The electronic equipment may be implemented to have achieved by any data recommendation device provided by the embodiment of the present invention
Effect is imitated, is detailed in the embodiment of front, details are not described herein.
The electronic equipment of the embodiment of the present invention first passes through according to DeepFM model and characteristic is preset, generates and update
Data generate target recommending data, are finally pushed to target recommending data further according to more new data and initial recommendation data
User improves the efficiency of data recommendation.
There is provided herein the various operations of embodiment.In one embodiment, one or more operations can be with structure
At the computer-readable instruction stored on one or more computer-readable mediums, will make to succeed in one's scheme when being executed by electronic equipment
It calculates equipment and executes the operation.Describing the sequences of some or all of operations, to should not be construed as to imply that these operations necessarily suitable
Sequence is relevant.It will be appreciated by those skilled in the art that the alternative sequence of the benefit with this specification.Furthermore, it is to be understood that
Not all operation must exist in each embodiment provided in this article.
Moreover, although the disclosure, this field skill has shown and described relative to one or more implementations
Art personnel will be appreciated that equivalent variations and modification based on the reading and understanding to the specification and drawings.The disclosure include it is all this
The modifications and variations of sample, and be limited only by the scope of the following claims.In particular, to by said modules (such as element,
Resource etc.) the various functions that execute, term for describing such components is intended to correspond to the specified function for executing the component
The random component (unless otherwise instructed) of energy (such as it is functionally of equal value), even if illustrated herein with execution in structure
The disclosure exemplary implementations in function open structure it is not equivalent.In addition, although the special characteristic of the disclosure
Through being disclosed relative to the only one in several implementations, but this feature can with such as can be to given or specific application
For be expectation and one or more other features combinations of other advantageous implementations.Moreover, with regard to term " includes ", " tool
Have ", " containing " or its deformation be used in specific embodiments or claims for, such term be intended to with term
The similar mode of "comprising" includes.
Each functional unit in the embodiment of the present invention can integrate in a processing module, be also possible to each unit list
It is solely physically present, can also be integrated in two or more units in a module.Above-mentioned integrated module can both use
Formal implementation of hardware can also be realized in the form of software function module.If the integrated module is with software function
The form of module is realized and when sold or used as an independent product, also can store in computer-readable storage Jie
In matter.Storage medium mentioned above can be read-only memory, disk or CD etc..Above-mentioned each device or system, can be with
Execute the method in correlation method embodiment.
Although the serial number before embodiment only makes for convenience of description in conclusion the present invention is disclosed above with embodiment
With not causing to limit to the sequence of various embodiments of the present invention.Also, above-described embodiment is not intended to limit the invention, this field
Those of ordinary skill, without departing from the spirit and scope of the present invention, can make it is various change and retouch, therefore it is of the invention
Protection scope subjects to the scope of the claims.
Claims (11)
1. a kind of data recommendation method characterized by comprising
The data recommendation request of user is obtained, and is requested according to the data recommendation, the corresponding initial recommendation of the user is obtained
Data and the corresponding characteristic of the user;
Data recommendation request is converted into the request of preset format, and according to the request of the preset format, is taken from model
It is engaged in obtaining default DeepFM model on device, wherein the request of the preset format meets the identification requirement of the model server;
According to the default DeepFM model and the characteristic, more new data is generated;
According to the more new data and the initial recommendation data, target recommending data is generated;
By the target recommending data, it is pushed to the user.
2. data recommendation method according to claim 1, which is characterized in that data recommendation request to be converted into presetting
The request of format, and according to the request of the preset format, default DeepFM model is obtained from model server, wherein described
The identification that the request of preset format meets the model server requires before step, further includes:
The history feature data of the user are obtained in real time;
Using the history feature data, the first DeepFM model is trained, generates multiple training results;
According to the multiple training result and the first DeepFM model, the default DeepFM model is generated, and by institute
Default DeepFM model is stated to be arranged on the model server.
3. data recommendation method according to claim 2, which is characterized in that described according to the training result and described
DeepFM model generates the default DeepFM model, and the default DeepFM model is arranged in the model server
Upper step, comprising:
According to the multiple training result and the first DeepFM model, multiple 2nd DeepFM models are generated;
Obtain the version information of the multiple 2nd DeepFM model;
According to the version information, from the multiple 2nd DeepFM model, the default DeepFM model is selected.
4. data recommendation method according to claim 1 to 3, which is characterized in that the model server includes
TFServing server, the request that data recommendation request is converted into preset format, and according to the preset format
Request, default DeepFM model is obtained from model server, wherein the request of the preset format meets the model clothes
The identification of business device requires the step to include:
Middleware services are used as using Proxy service, data recommendation request is converted into the request of gRPC format;
By the request of the gRPC format, from the TFServing server, the default DeepFM model is obtained.
5. data recommendation method according to claim 1 to 3, which is characterized in that the default DeepFM model
It is described according to the default DeepFM model and the characteristic including FM model and DNN model, generate more new data step
Suddenly, comprising:
Using the FM model, by the characteristic vectorization;
By the characteristic of vectorization, the DNN model is inputted, generates the first output result;
According to first output as a result, and the FM model corresponding second exports as a result, more new data described in generating.
6. a kind of data recommendation device characterized by comprising
First obtains module, and the data recommendation for obtaining user is requested, and is requested according to the data recommendation, obtains the use
The corresponding initial recommendation data in family and the corresponding characteristic of the user;
Conversion module, for data recommendation request to be converted into the request of preset format, and according to the preset format
Request obtains default DeepFM model, wherein the request of the preset format meets the model service from model server
The identification requirement of device;
First generation module, for generating more new data according to the default DeepFM model and the characteristic;
Second generation module, for generating target recommending data according to the more new data and the initial recommendation data;
Pushing module, for being pushed to the user for the target recommending data.
7. data recommendation device according to claim 6, which is characterized in that the data recommendation device further include:
Second obtains module, for obtaining the history feature data of the user in real time;
Training module is trained the first DeepFM model, generates multiple training knots for using the history feature data
Fruit;
Generation module, for generating described default according to the multiple training result and the first DeepFM model
DeepFM model, and the default DeepFM model is arranged on the model server.
8. data recommendation device according to claim 7, which is characterized in that the generation module includes:
Submodule is generated, for generating multiple second according to the multiple training result and the first DeepFM model
DeepFM model;
First acquisition submodule, for obtaining the version information of the multiple 2nd DeepFM model;
Submodule is chosen, for from the multiple 2nd DeepFM model, selecting described default according to the version information
DeepFM model.
9. according to data recommendation device described in claim 6-8 any one, which is characterized in that the conversion module includes:
Data recommendation request is converted into gRPC lattice for being used as middleware services using Proxy service by transform subblock
The request of formula;
Second acquisition submodule from the TFServing server, obtains institute for the request by the gRPC format
State default DeepFM model.
10. according to data recommendation device described in claim 6-8 any one, which is characterized in that the default DeepFM mould
Type includes FM model and DNN model, and first generation module includes:
Vectorization submodule, for using the FM model, by the characteristic vectorization;
Input submodule generates the first output result for inputting the DNN model for the characteristic of vectorization;
Submodule is updated, for being exported according to described first as a result, and the FM model corresponding second exports as a result, generation
The more new data.
11. a kind of storage medium, is stored with processor-executable instruction, which is provided such as by executing described instruction
Any data recommendation method in claim 1-5.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810670441.0A CN108989397B (en) | 2018-06-26 | 2018-06-26 | Data recommendation method and device and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810670441.0A CN108989397B (en) | 2018-06-26 | 2018-06-26 | Data recommendation method and device and storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108989397A true CN108989397A (en) | 2018-12-11 |
CN108989397B CN108989397B (en) | 2021-04-20 |
Family
ID=64538770
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810670441.0A Active CN108989397B (en) | 2018-06-26 | 2018-06-26 | Data recommendation method and device and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108989397B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109902224A (en) * | 2019-01-17 | 2019-06-18 | 平安城市建设科技(深圳)有限公司 | Source of houses recommended method, device, equipment and medium based on user behavior analysis |
CN110166522A (en) * | 2019-04-01 | 2019-08-23 | 腾讯科技(深圳)有限公司 | Server identification method, device, readable storage medium storing program for executing and computer equipment |
CN110457589A (en) * | 2019-08-19 | 2019-11-15 | 上海新共赢信息科技有限公司 | A kind of vehicle recommended method, device, equipment and storage medium |
CN110674416A (en) * | 2019-09-20 | 2020-01-10 | 北京小米移动软件有限公司 | Game recommendation method and device |
CN111368150A (en) * | 2018-12-25 | 2020-07-03 | 北京奇虎科技有限公司 | Training method and device of prediction model and electronic equipment |
CN111859169A (en) * | 2019-06-05 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | Destination recommendation method and system |
CN112100362A (en) * | 2020-09-11 | 2020-12-18 | 北京百度网讯科技有限公司 | Document format recommendation model training method and device and electronic equipment |
CN112287239A (en) * | 2020-12-30 | 2021-01-29 | 平安科技(深圳)有限公司 | Course recommendation method and device, electronic equipment and storage medium |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080208786A1 (en) * | 2007-02-22 | 2008-08-28 | Serrano-Morales Carlos A | Rule-based management of adaptive models and agents |
CN101321190A (en) * | 2008-07-04 | 2008-12-10 | 清华大学 | Recommend method and recommend system of heterogeneous network |
US20090281923A1 (en) * | 2008-05-06 | 2009-11-12 | David Selinger | System and process for improving product recommendations for use in providing personalized advertisements to retail customers |
CN103678647A (en) * | 2013-12-20 | 2014-03-26 | Tcl集团股份有限公司 | Method and system for recommending information |
CN104216886A (en) * | 2013-05-29 | 2014-12-17 | 酷盛(天津)科技有限公司 | Video recommendation device, system and method |
CN106874522A (en) * | 2017-03-29 | 2017-06-20 | 珠海习悦信息技术有限公司 | Information recommendation method, device, storage medium and processor |
CN107911491A (en) * | 2017-12-27 | 2018-04-13 | 广东欧珀移动通信有限公司 | Information recommendation method, device and storage medium, server and mobile terminal |
-
2018
- 2018-06-26 CN CN201810670441.0A patent/CN108989397B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080208786A1 (en) * | 2007-02-22 | 2008-08-28 | Serrano-Morales Carlos A | Rule-based management of adaptive models and agents |
US20090281923A1 (en) * | 2008-05-06 | 2009-11-12 | David Selinger | System and process for improving product recommendations for use in providing personalized advertisements to retail customers |
CN101321190A (en) * | 2008-07-04 | 2008-12-10 | 清华大学 | Recommend method and recommend system of heterogeneous network |
CN104216886A (en) * | 2013-05-29 | 2014-12-17 | 酷盛(天津)科技有限公司 | Video recommendation device, system and method |
CN103678647A (en) * | 2013-12-20 | 2014-03-26 | Tcl集团股份有限公司 | Method and system for recommending information |
CN106874522A (en) * | 2017-03-29 | 2017-06-20 | 珠海习悦信息技术有限公司 | Information recommendation method, device, storage medium and processor |
CN107911491A (en) * | 2017-12-27 | 2018-04-13 | 广东欧珀移动通信有限公司 | Information recommendation method, device and storage medium, server and mobile terminal |
Non-Patent Citations (3)
Title |
---|
DANIEL GALLEGO等: ""A model for generating proactive context-aware recommendations in e-Learning systems"", 《IEEE》 * |
文哥的学习日记: ""推荐系统遇上深度学习(三)--DeepFM模型理论和实践"", 《简书》 * |
陈亮等: ""基于DNN算法的移动视频推荐策略"", 《CNKI》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111368150A (en) * | 2018-12-25 | 2020-07-03 | 北京奇虎科技有限公司 | Training method and device of prediction model and electronic equipment |
CN111368150B (en) * | 2018-12-25 | 2024-04-02 | 北京奇虎科技有限公司 | Training method and device of prediction model and electronic equipment |
CN109902224A (en) * | 2019-01-17 | 2019-06-18 | 平安城市建设科技(深圳)有限公司 | Source of houses recommended method, device, equipment and medium based on user behavior analysis |
CN110166522B (en) * | 2019-04-01 | 2021-08-24 | 腾讯科技(深圳)有限公司 | Server identification method and device, readable storage medium and computer equipment |
CN110166522A (en) * | 2019-04-01 | 2019-08-23 | 腾讯科技(深圳)有限公司 | Server identification method, device, readable storage medium storing program for executing and computer equipment |
CN111859169A (en) * | 2019-06-05 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | Destination recommendation method and system |
CN110457589A (en) * | 2019-08-19 | 2019-11-15 | 上海新共赢信息科技有限公司 | A kind of vehicle recommended method, device, equipment and storage medium |
CN110457589B (en) * | 2019-08-19 | 2020-05-12 | 上海新共赢信息科技有限公司 | Vehicle recommendation method, device, equipment and storage medium |
CN110674416A (en) * | 2019-09-20 | 2020-01-10 | 北京小米移动软件有限公司 | Game recommendation method and device |
CN112100362B (en) * | 2020-09-11 | 2023-07-28 | 北京百度网讯科技有限公司 | Document format recommendation model training method and device and electronic equipment |
CN112100362A (en) * | 2020-09-11 | 2020-12-18 | 北京百度网讯科技有限公司 | Document format recommendation model training method and device and electronic equipment |
CN112287239B (en) * | 2020-12-30 | 2021-03-19 | 平安科技(深圳)有限公司 | Course recommendation method and device, electronic equipment and storage medium |
CN112287239A (en) * | 2020-12-30 | 2021-01-29 | 平安科技(深圳)有限公司 | Course recommendation method and device, electronic equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN108989397B (en) | 2021-04-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108989397A (en) | Data recommendation method, device and storage medium | |
CN107741976B (en) | Intelligent response method, device, medium and electronic equipment | |
CN111079015B (en) | Recommendation method and device, computer equipment and storage medium | |
CN111090686B (en) | Data processing method, device, server and storage medium | |
CN111209477B (en) | Information recommendation method and device, electronic equipment and storage medium | |
CN111209309A (en) | Method, device and equipment for determining processing result of data flow graph and storage medium | |
CN108491408A (en) | A kind of processing method of action message, device, electronic equipment and storage medium | |
CN110399564B (en) | Account classification method and device, storage medium and electronic device | |
CN110334198A (en) | Intelligent robot method of servicing, system and terminal device | |
CN109409419A (en) | Method and apparatus for handling data | |
CN117273018A (en) | Information processing method, information processing apparatus, computer device, and storage medium | |
CN112036987A (en) | Method and device for determining recommended commodities | |
CN115563160A (en) | Data processing method, data processing device, computer equipment and computer readable storage medium | |
CN110471968A (en) | Dissemination method, device, equipment and the storage medium of ETL task | |
CN112035581B (en) | Model-based task processing method, device, equipment and medium | |
CN114677168A (en) | Resource recommendation method, device, equipment and medium | |
CN114387066A (en) | Commodity information transmission method, system, electronic device and storage medium | |
CN108804442A (en) | Sequence number generation method and device | |
CN113283517A (en) | Household service matching method, system and computer equipment | |
CN108874625A (en) | Information processing method and device, electronic equipment, storage medium | |
CN117453988A (en) | Product recommendation method and device | |
CN113568936B (en) | Real-time stream data storage method, device and terminal equipment | |
CN113779369B (en) | Matching method, matching device, electronic equipment and storage medium | |
CN115964556A (en) | Content recommendation method, device, computer equipment and computer readable storage medium | |
CN114003761A (en) | Video interaction method, device, system, equipment and computer readable storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |