CN113672820A - Training method of feature extraction network, information recommendation method, device and equipment - Google Patents

Training method of feature extraction network, information recommendation method, device and equipment Download PDF

Info

Publication number
CN113672820A
CN113672820A CN202110902980.4A CN202110902980A CN113672820A CN 113672820 A CN113672820 A CN 113672820A CN 202110902980 A CN202110902980 A CN 202110902980A CN 113672820 A CN113672820 A CN 113672820A
Authority
CN
China
Prior art keywords
information
historical
recommendation information
recommended
behavior
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
Application number
CN202110902980.4A
Other languages
Chinese (zh)
Other versions
CN113672820B (en
Inventor
彭冲
程兵
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Sankuai Online Technology Co Ltd
Original Assignee
Beijing Sankuai Online Technology Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beijing Sankuai Online Technology Co Ltd filed Critical Beijing Sankuai Online Technology Co Ltd
Priority to CN202110902980.4A priority Critical patent/CN113672820B/en
Publication of CN113672820A publication Critical patent/CN113672820A/en
Application granted granted Critical
Publication of CN113672820B publication Critical patent/CN113672820B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9536—Search customisation based on social or collaborative filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53—Querying
    • G06F16/535—Filtering based on additional data, e.g. user or group profiles
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • 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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Abstract

The application discloses a training method, an information recommendation method, a device and equipment of a feature extraction network, and belongs to the technical field of computers. The method comprises the following steps: acquiring exposure recommendation information and first historical behavior information; acquiring a historical recommendation information sequence; acquiring a second historical behavior information sequence according to the historical recommendation information sequence; training a feature extraction network through the exposure recommendation information, the first historical behavior information, the historical recommendation information sequence and the second historical behavior information sequence, wherein the feature extraction network is used for extracting semantic vectors of the exposure recommendation information and the historical recommendation information. The semantic vector of the information can be extracted through the feature extraction network, so that the information which is interested can be recommended to a user based on the matching result of the semantic vector. In the training process, the used training data are data in an actual application scene, so that the accuracy of the extracted semantic vector can be improved, and the accuracy when the image-text combination is recommended to the user is improved.

Description

Training method of feature extraction network, information recommendation method, device and equipment
Technical Field
The present application relates to the field of computer technologies, and in particular, to a training method, an information recommendation apparatus, and a device for a feature extraction network.
Background
The client with the recommendation function can recommend a picture and text combination which is interesting to the user, wherein the picture and text combination is a document (item) formed by pictures and titles, and the titles are used for describing the pictures.
In the process of recommending the image-text combination, the server needs to extract a first semantic vector of the image-text combination to be recommended through a machine learning model, then determines the image-text combination to be recommended matched with the image-text combination which is interested by the user in the past according to the similarity between a second semantic vector of the image-text combination which is interested by the user in the past and the first semantic vector, and recommends the image-text combination to the user, so that the image-text combination which is interested by the user is recommended to the user.
When the image-text combination is recommended to the user in the mode, the machine learning model needs to be pre-trained by using sample data. However, the sample data has wide sources and may have great difference with data in an actual application scene, which causes inaccuracy of semantic vectors of extracted image-text combinations, and thus causes lower accuracy of recommended image-text combinations.
Disclosure of Invention
The application provides a training method, an information recommendation device and equipment for a feature extraction network, which can improve the accuracy of recommending image-text combination. The technical scheme is as follows:
according to an aspect of the present application, there is provided a training method of a feature extraction network, the method including:
acquiring exposure recommendation information and first historical behavior information, wherein the exposure recommendation information is recommended to a sample user account in a sample user account set, and the first historical behavior information is used for reflecting that the sample user account and the exposure recommendation information generate a first interactive behavior;
acquiring a historical recommendation information sequence, wherein the historical recommendation information sequence comprises historical recommendation information recommended to the sample user account before the exposure recommendation information;
acquiring a second historical behavior information sequence according to the historical recommendation information sequence, wherein the second historical behavior information sequence comprises second historical behavior information which is used for reflecting that the sample user account and the historical recommendation information generate a second interaction behavior;
training a feature extraction network through the exposure recommendation information, the first historical behavior information, the historical recommendation information sequence and the second historical behavior information sequence, wherein the feature extraction network is used for extracting semantic vectors of the exposure recommendation information and the historical recommendation information.
According to another aspect of the present application, there is provided an information recommendation method applied in a computer device running a feature extraction network, the feature extraction network being trained by the training method of the feature extraction network as described above, the method including:
determining reference to-be-recommended information corresponding to a user account to be recommended, wherein the reference to-be-recommended information comprises at least one of information to be recommended which generates interactive behaviors with the user account to be recommended and information to be recommended which generates interactive behaviors with the user account matched with the characteristics of the user account to be recommended;
extracting a third semantic vector of the reference information to be recommended through the feature extraction network;
obtaining a fourth semantic vector of a plurality of candidate information to be recommended, wherein the fourth semantic vector is obtained by extracting the candidate information to be recommended through the feature extraction network;
and determining a target fourth semantic vector matched with the third semantic vector in fourth semantic vectors of the candidate information to be recommended, and determining the candidate information to be recommended corresponding to the target fourth semantic vector as recommendation information, wherein the recommendation information is used for recommending to the user account to be recommended.
According to another aspect of the present application, there is provided a training apparatus for a feature extraction network, the apparatus including:
the system comprises an acquisition module, a processing module and a display module, wherein the acquisition module is used for acquiring exposure recommendation information and first historical behavior information, the exposure recommendation information is recommended to a sample user account in a sample user account set, and the first historical behavior information is used for reflecting that the sample user account and the exposure recommendation information generate a first interaction behavior;
the acquisition module is further configured to acquire a historical recommendation information sequence, where the historical recommendation information sequence includes historical recommendation information recommended to the sample user account before the exposure recommendation information;
the obtaining module is further configured to obtain a second historical behavior information sequence according to the historical recommendation information sequence, where the second historical behavior information sequence includes second historical behavior information, and the second historical behavior information is used to reflect that the sample user account and the historical recommendation information generate a second interaction behavior;
and the training module is used for training a feature extraction network through the exposure recommendation information, the first historical behavior information, the historical recommendation information sequence and the second historical behavior information sequence, wherein the feature extraction network is used for extracting the exposure recommendation information and semantic vectors of the historical recommendation information.
In an alternative design, the feature extraction network is a sub-network in a machine learning model that further includes a behavior prediction network in cascade with the feature extraction network; the training module is configured to:
extracting a first semantic vector of the exposure recommendation information and a second semantic vector of each history recommendation information in the history recommendation information sequence through the feature extraction network;
fusing a second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the characteristics of second historical behavior information corresponding to the historical recommendation information to obtain fused characteristics corresponding to the historical recommendation information;
predicting predicted behavior information of the exposure recommendation information through the behavior prediction network based on the first semantic vector and fusion features corresponding to each piece of historical recommendation information in the historical recommendation information sequence;
training the machine learning model according to a difference between the first historical behavior information and the predicted behavior information.
In an alternative design, the machine learning model further includes a behavior encoding network, the behavior encoding network being cascaded with the feature extraction network and the behavior prediction network; the training module is configured to:
processing each second historical behavior information in the second historical behavior information sequence through the behavior coding network to obtain a historical behavior coding vector corresponding to the second historical behavior information;
and fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the historical behavior coding vector of the second piece of historical behavior information corresponding to the historical recommendation information to obtain the fusion characteristic.
In an optional design, the machine learning model further includes a behavior fusion network, and the behavior fusion network is cascaded with the behavior prediction network, the behavior coding network, and the feature extraction network; the training module is used for
And fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the historical behavior coding vector of the second piece of historical behavior information corresponding to the historical recommendation information through the behavior fusion network to obtain the fusion characteristic.
In an optional design, the content of the exposure recommendation information and the history recommendation information is a text-text combination, the text-text combination is a document composed of a picture and a title, and the title is used for describing the picture.
In an alternative design, the obtaining module is configured to:
acquiring at least one of a browsing information sequence and an interactive information sequence;
the browsing information sequence is composed of browsing recommendation information recommended to the sample user account before the recommendation time of the exposure recommendation information, and the interaction information sequence is composed of interaction recommendation information generating a third interaction behavior with the sample user account before the recommendation time of the exposure recommendation information.
In an optional design, the interactive recommendation information includes at least one of clicked recommendation information, praised recommendation information, shared recommendation information, collected recommendation information, and commented recommendation information.
According to another aspect of the present application, there is provided an information recommendation apparatus operating with a feature extraction network trained by a training apparatus of the feature extraction network as described above, the apparatus including:
the system comprises a determining module, a recommending module and a recommending module, wherein the determining module is used for determining reference to-be-recommended information corresponding to a user account to be recommended, and the reference to-be-recommended information comprises at least one of information to be recommended which generates an interactive behavior with the user account to be recommended and is matched with the characteristics of the user account to be recommended;
the extraction module is used for extracting a third semantic vector of the reference information to be recommended through the feature extraction network;
the acquisition module is used for acquiring a fourth semantic vector of a plurality of candidate information to be recommended, and the fourth semantic vector is obtained by extracting the candidate information to be recommended through the feature extraction network;
the determining module is further configured to determine, in a fourth semantic vector of the multiple candidate information to be recommended, a target fourth semantic vector matched with the third semantic vector, and determine candidate information to be recommended corresponding to the target fourth semantic vector as recommendation information, where the recommendation information is used for recommending to the user account to be recommended.
In an optional design, the content of the reference information to be recommended and the candidate information to be recommended is a text-text combination, the text-text combination is a document composed of a picture and a title, and the title is used for describing the picture.
According to another aspect of the present application, there is provided a computer device comprising a processor and a memory, the memory having stored therein at least one instruction, at least one program, a set of codes, or a set of instructions, which is loaded and executed by the processor to implement the training method or the information recommendation method of the feature extraction network as described above.
According to another aspect of the present application, there is provided a computer-readable storage medium having stored therein at least one instruction, at least one program, code set, or set of instructions that is loaded and executed by a processor to implement the training method or information recommendation method of the feature extraction network as described above.
According to another aspect of the application, a computer program product or computer program is provided, comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to execute the training method or the information recommendation method of the feature extraction network provided in the various alternative implementations of the above aspects.
The beneficial effect that technical scheme that this application provided brought includes at least:
the semantic vector of the information can be extracted through the trained feature extraction network, so that the interested information can be recommended to the user based on the matching result of the semantic vector. In the process of training the feature extraction network, the used training data is data in an actual application scene, so that the accuracy of the extracted semantic vector can be improved, the accuracy when recommending information to a user is improved, and the accuracy when recommending image-text combination to the user can be improved.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a schematic diagram of a process for training a machine learning model provided by an exemplary embodiment of the present application;
FIG. 2 is a schematic flow chart diagram of a training method for a feature extraction network provided in an exemplary embodiment of the present application;
FIG. 3 is a schematic flow chart diagram illustrating a method for training a feature extraction network according to another exemplary embodiment of the present application;
FIG. 4 is a flowchart illustrating an information recommendation method according to an exemplary embodiment of the present application;
FIG. 5 is a diagram of a recommended teletext combination provided in an exemplary embodiment of the application;
FIG. 6 is a schematic structural diagram of a training apparatus for a feature extraction network according to an exemplary embodiment of the present application;
fig. 7 is a schematic structural diagram of an information recommendation device according to an exemplary embodiment of the present application;
fig. 8 is a schematic structural diagram of a computer device according to an exemplary embodiment of the present application.
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and together with the description, serve to explain the principles of the application.
Detailed Description
To make the objects, technical solutions and advantages of the present application more clear, embodiments of the present application will be described in further detail below with reference to the accompanying drawings.
First, terms referred to in the embodiments of the present application are described:
combining the pictures and texts: a teletext combination is a document made up of pictures and titles, where a title is used to describe a picture. For example, a teletext combination is made up of a picture of an animal and a title describing the picture of the animal.
Exposure and image-text combination: the exposure teletext combination is the combination of teletext already recommended to the user account. For example, the image-text combination is displayed in the client terminal where the user account is logged in.
Browsing image-text combination: the browsing image-text combination refers to the image-text combination browsed by the user account before the recommending time of recommending the exposure image-text combination to the user account. For example, the combination of graphics and text displayed in the client that the user account logs in before the recommended time.
Interactive image-text combination: the interactive image-text combination refers to the image-text combination which has interactive behaviors with the user account before the recommendation time of the exposure image-text combination to the user account. The interactive behavior comprises at least one of click, praise, share, favorite and comment. For example, the user account has approved a combination of text and text on the client that the user logged in before the recommendation time.
Exposure history behavior information: and the exposure history behavior information is used for reflecting the interactive behavior generated by the combination of the user account and the exposure graphics and text. The exposure history behavior information is generated under the condition that the user account receiving the exposure graphics context combination and the exposure graphics context combination generate the interactive behavior. For example, the exposure history behavior information can reflect which interactive behaviors exist between the user account and the exposure-graphics combination and the time when the interactive behaviors are generated.
Browsing historical behavior information: and the browsing history behavior information is used for reflecting the interactive behavior generated by the combination of the user account and the browsing image-text. The browsing history behavior information is generated under the condition that the user account which is combined by browsing the pictures and texts and the browsing pictures and texts generates the interaction behavior. For example, the browsing history behavior information can reflect which interaction behaviors exist between the user account and the browsing image-text combination and the time when the interaction behaviors are generated.
Interaction history behavior information: and the interactive historical behavior information is used for reflecting the interactive behavior generated by the interactive graphics and text combination of the user account. For example, the interactive historical behavior information can reflect which interactive behaviors exist between the user account and the interactive graphics and text combination and the time when the interactive behaviors are generated.
Candidate image-text combination: the candidate teletext combinations are those used for recommending to the user account.
FIG. 1 is a schematic diagram of a process for training a machine learning model provided by an exemplary embodiment of the present application. As shown in fig. 1, the machine learning model includes a feature extraction network 101, a behavior encoding network 102, a behavior fusion network 103, and a behavior prediction network 104.
When training the machine learning model, the computer device obtains an exposure image-text combination 105, a browsing image-text combination sequence 106, an interactive image-text combination sequence 107, exposure history behavior information corresponding to the exposure image-text combination 105, browsing history behavior information 108 corresponding to the browsing image-text combination sequence 106, and interactive history behavior information 109 corresponding to the interactive image-text combination sequence 107. Where the exposure teletext combination 105 is a teletext combination recommended to a sample user account in the set of sample user accounts. The browsing teletext combination sequence 106 is formed by the teletext combinations browsed by the sample user account before the recommended moment of exposure teletext combination 105. The interactive teletext sequence 107 is formed by teletext combinations that have an interactive activity with the sample user account before the recommended time of the exposure teletext combination 105. The browsing history behavior information is used to reflect the interaction behavior generated by the sample user account and exposure text combination 105. The browsing history behavior information 108 is used to reflect that the interaction behavior generated by the sample user account and the teletext combination in the browsing teletext combination sequence 106. The interactive historical behavior information 109 is used to reflect that the interactive behavior generated by the image-text combination in the sequence 107 of the interactive image-text combination with the sample user account number.
After acquiring the training data, the computer device extracts an exposure-graphics-text semantic vector 110 of the exposure-graphics-text combination 105 through the feature extraction network 101, extracts a browsing-graphics-text semantic vector 111 of the graphics-text combination in the browsing-graphics-text combination sequence 106 through the feature extraction network 101, and extracts an interactive-graphics-text semantic vector 112 of the graphics-text combination in the interactive-graphics-text combination sequence 107 through the feature extraction network 101. Furthermore, the computer device processes the browsing history behavior information 108 through the behavior coding network 102 to obtain a browsing image-text behavior coding vector 113, and processes the interactive history behavior information 109 through the behavior coding network 102 to obtain an interactive image-text behavior coding vector 114. Then the computer equipment performs feature fusion on the browsing image-text semantic vector 111 and the corresponding browsing image-text behavior coding vector 113 through the behavior fusion network 103 to obtain a browsing image-text fusion vector 115, and performs feature fusion on the interactive image-text semantic vector 112 and the corresponding interactive image-text behavior coding vector 114 through the behavior fusion network 103 to obtain an interactive image-text fusion vector 116. So that the semantic vector of each image-text combination is superimposed with the behavior information of the interaction behavior corresponding to the image-text combination. The computer device then predicts the predicted behavior information 117 of the exposure-teletext combination 105 over the behavior prediction network based on the exposure-teletext semantic vector 110, the browsing-teletext fusion vector 115 and the interactive-teletext fusion vector 116, and trains the machine learning model based on the difference between the predicted behavior information 117 and the exposure-history behavior information of the exposure-teletext combination 105. The machine learning model can predict the interaction behavior possibly generated by the user account and a certain image-text combination according to the image-text combination browsed or interacted by the user account before browsing the certain image-text combination and the historical behavior information generated by the interaction behavior of the browsed or interacted image-text combination.
Through the feature extraction network 101 of the trained machine learning model, the computer equipment can extract semantic vectors of the candidate image-text combination and match the semantic vectors of the candidate image-text combination with semantic vectors of image-text combinations which are interested by the account of the user to be recommended, so that the recommended image-text combination for recommending the account of the user to be recommended is determined in the candidate image-text combination. The image-text combination which is interested by the user account to be recommended comprises an image-text combination which generates interactive behaviors with the user account to be recommended and has characteristics matched with (similar to) the characteristics of the user account to be recommended.
The semantic vectors of the candidate image-text combinations and the semantic vectors of the image-text combinations which are interesting to the user in the past are extracted through the feature extraction network, and the recommended image-text combinations can be determined in the candidate image-text combinations based on the matching results of the semantic vectors, so that the interesting image-text combinations are recommended to the user. In the process of training the feature extraction network, the used training data is data in an actual application scene, so that the accuracy of the semantic vector of the extracted image-text combination can be improved, and the accuracy of the recommended image-text combination is improved.
In addition, in the process of training the machine learning model, manual sample labeling is not needed (if data in an actual application scene is directly used, manual sample labeling is usually needed), but the machine learning model is trained based on the processing result of the existing data, so that the efficiency of recommending the image-text combination can be improved.
In addition, the machine learning model for extracting the semantic vector of the image-text combination can learn the relationship between the characteristics of a certain image-text combination browsed by a user, the characteristics of the image-text combination browsed or interacted by the user before the certain image-text combination, the characteristics of the historical behavior information of the browsed or interacted image-text combination and the interaction behavior generated by the user, so that the machine learning model can extract the characteristics related to the behavior (interest) of the user in the image-text combination, and the accuracy of subsequent recommendation of image-text combination is improved.
Fig. 2 is a flowchart illustrating a training method of a feature extraction network according to an exemplary embodiment of the present application. The method may be used in a computer device. As shown in fig. 2, the method includes:
step 202: exposure recommendation information and first historical behavior information are obtained.
The exposure recommendation information is recommendation information recommended to a sample user account in the sample user account set. The sample set of user accounts is comprised of sample user accounts that received the exposure recommendation information. The exposure recommendation information and the first historical behavior information are information generated in an actual application scene. Illustratively, the exposure recommendation information is information displayed in the client to which the sample user account is logged in, corresponding to the computer device.
The client corresponds to a computer device, and the client has a function of recommending information to a user, for example, the client can be a local life service client. The computer device can be a server, which is a server, or a server cluster composed of several servers, or a virtual server in a cloud computing service center, etc.
Optionally, the content of the exposure recommendation information is a text-text combination, and the text-text combination is a document composed of a picture and a title, where the title is used for describing the picture. For example, the teletext combination is composed of a picture of a cat and the title "this cat really loves".
The first historical behavior information is used for reflecting that the sample user account and the exposure recommendation information generate a first interactive behavior. The first interactive behavior includes at least one of click, like, share, favorite, and comment. That is, the first historical behavior information can reflect at least one of whether the exposure recommendation information is clicked, whether the exposure recommendation information is praised, whether the exposure recommendation information is shared, whether the exposure recommendation information is collected, and whether the exposure recommendation information is commented.
Step 204: and acquiring a historical recommendation information sequence.
The sequence of historical recommendation information includes historical recommendation information recommended to the sample user account prior to exposing the recommendation information. In the historical recommendation information sequence, the historical recommendation information is sorted according to the recommendation time for recommending to the sample user account. The historical recommendation information is information generated in an actual application scenario. Optionally, the content of the history recommendation information is a graph-text combination, where the graph-text combination is a document composed of a picture and a title, and the title is used for describing the picture.
Optionally, the historical recommendation information sequence includes at least one of a browsing information sequence and an interaction information sequence. The browsing information sequence is composed of browsing recommendation information recommended to the sample user account before the recommendation time of the exposure recommendation information, and the interactive information sequence is composed of interactive recommendation information generating a third interactive behavior with the sample user account before the recommendation time of the exposure recommendation information. The third interactive behavior includes at least one of click, like, share, favorite, and comment.
The computer device can further divide the interaction information sequence according to different third interaction behaviors. For example, the interactive recommendation information with click behavior is divided into click information sequences. And dividing the interactive recommendation information of the behavior of the click approval, sharing and comment into other interactive information sequences. It should be noted that there is a coincidence between the information in the browsing information sequence and the information in the interactive information sequence. Further, for each of the divided sequences in the mutual information sequence, there is also a coincidence between the information in each sequence.
Optionally, the computer device may determine the historical recommendation information using at least one of a time period limit or a number limit in obtaining the sequence of historical recommendation information. The time interval limitation means that the computer device determines a history recommendation information sequence based on the history recommendation information in the history time interval before the recommendation time of the exposure recommendation information. The quantity limit means that the computer device determines a history recommendation information sequence based on history recommendation information of which the quantity before the recommendation time of the exposure recommendation information is smaller than a quantity threshold. For example, the computer device is at 11:00 recommends exposure recommendation information to the account number of the sample user, and when the historical recommendation sequence is obtained, the computer equipment only obtains the historical recommendation information recommended to the account number of the sample user between 9:00 and 11:00 to obtain the historical recommendation information sequence.
Step 206: and acquiring a second historical behavior information sequence according to the historical recommendation information sequence.
The second historical behavior information sequence comprises second historical behavior information, and the second historical behavior information is used for reflecting that the sample user account and the historical recommendation information generate second interaction behaviors. I.e. the second historical behavior information is generated based on the interaction between the sample user account and the historical recommendation information. The second historical behavior information is information generated in an actual application scenario. Optionally, the second interactive behavior includes at least one of click, like, share, favorite, and comment.
Step 208: and training the feature extraction network through the exposure recommendation information, the first historical behavior information, the historical recommendation information sequence and the second historical behavior information sequence.
The feature extraction network is used for extracting semantic vectors of exposure recommendation information and historical recommendation information. Through the feature extraction network, semantic vectors of reference recommendation information and semantic vectors of candidate recommendation information which are interested in the past of the user account can be extracted. By matching the semantic vector of the reference recommendation information which is interested in the user account in the past with the semantic vector of the candidate recommendation information, the recommendation information can be determined from the candidate recommendation information and recommended to the user account.
Optionally, the feature extraction network can characterize a model for a deformation-based bi-directional Encoder Image (Image Bidirectional Encoder Representations from transformations, ImageBERT). The feature extraction network belongs to a machine learning model. The computer device can predict the predicted behavior information of the exposure recommendation information through the machine learning model based on the exposure recommendation information, the historical recommendation information sequence, and the second historical behavior information. The predicted behavior information is used to reflect which interaction behaviors may be generated by the sample user account and the exposure recommendation information. And then, according to the difference between the predicted behavior information and the first historical behavior information, training the machine learning model can be realized. The feature extraction network is used for extracting semantic vectors of exposure recommendation information and historical recommendation information in the machine learning model and taking the semantic vectors as input when behavior information is predicted subsequently.
In summary, in the method provided in this embodiment, the semantic vector of the information can be extracted through the feature extraction network obtained through training, so that the information of interest can be recommended to the user based on the matching result of the semantic vector. In the process of training the feature extraction network, the used training data is data in an actual application scene, so that the accuracy of the extracted semantic vector can be improved, the accuracy when recommending information to a user is improved, and the accuracy when recommending image-text combination to the user can be improved.
Fig. 3 is a flowchart illustrating a training method for a feature extraction network according to another exemplary embodiment of the present application. The method may be used in a computer device. As shown in fig. 3, the method includes:
step 302: exposure recommendation information and first historical behavior information are obtained.
The exposure recommendation information is recommendation information recommended to a sample user account in the sample user account set. The first historical behavior information is used for reflecting that the sample user account and the exposure recommendation information generate a first interactive behavior. Optionally, the first interactive behavior comprises at least one of click, like, share, favorite, and comment. The exposure recommendation information and the first historical behavior information are information generated in an actual application scene.
Step 304: and acquiring a historical recommendation information sequence.
The historical recommendation information sequence includes historical recommendation information recommended to the sample user account prior to exposing the recommendation information. In the historical recommendation information sequence, the historical recommendation information is sorted according to the recommendation time for recommending to the sample user account. The historical recommendation information is information generated in an actual application scenario.
Optionally, the computer device obtains the historical recommendation information sequence, including obtaining at least one of a browsing information sequence and an interaction information sequence. The browsing information sequence is composed of browsing recommendation information recommended to the sample user account before the recommendation time of the exposure recommendation information. The interactive information sequence is composed of interactive recommendation information which generates a third interactive behavior with the sample user account before the recommendation time of the exposure recommendation information.
The third interactive behavior includes at least one of click, like, share, favorite, and comment. The interactive recommendation information includes at least one of clicked recommendation information, approved recommendation information, shared recommendation information, collected recommendation information and commented recommendation information.
The computer device can further divide the interaction information sequence according to different third interaction behaviors. For example, the interactive recommendation information with click behavior is divided into click information sequences. And dividing the interactive recommendation information of the behavior of the click approval, sharing and comment into other interactive information sequences.
The contents of the exposure recommendation information and the history recommendation information may be a combination of text and text. A teletext combination is a document made up of pictures and titles, which are used to describe the pictures.
Step 306: and acquiring a second historical behavior information sequence according to the historical recommendation information sequence.
The second historical behavior information sequence comprises second historical behavior information, and the second historical behavior information is used for reflecting that the sample user account and the historical recommendation information generate second interaction behaviors. I.e. the second historical behavior information is generated based on the interaction between the sample user account and the historical recommendation information. The second historical behavior information is information generated in an actual application scenario. Optionally, the second interactive behavior includes at least one of click, like, share, favorite, and comment.
Step 308: and extracting a first semantic vector of the exposure recommendation information and a second semantic vector of each history recommendation information in the history recommendation information sequence through a feature extraction network.
The first semantic vector is used for reflecting the characteristics of the exposure recommendation information. When the content of the exposure recommendation information is the image-text combination, the first semantic vector can reflect the characteristics of the title and the image in the exposure recommendation information at the same time. The second semantic vector is used for reflecting the characteristics of the historical recommendation information. When the content of the historical recommendation information is image-text combination, the second semantic vector can reflect the characteristics of the title and the picture in the historical recommendation information at the same time.
It should be noted that different features may be generated by the same picture and different combinations of titles, and different features may be generated by different pictures and the same combinations of titles.
The feature extraction network is based on Neural Networks (NN), and can realize multi-modal calculation of a first semantic vector of exposure recommendation information and a second semantic vector of each history recommendation information in a history recommendation information sequence. Under the condition that the contents of the exposure recommendation information and the history recommendation information are image-text combinations, the feature extraction network respectively extracts the features of the images in the image-text combinations and the features of the titles and then performs fusion. Optionally, the feature extraction network can be ImageBERT.
Step 310: and fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the characteristics of the second piece of historical behavior information corresponding to the historical recommendation information to obtain fused characteristics corresponding to the historical recommendation information.
In the process of fusing the features of the second semantic vector and the second historical behavior information, the computer device fuses the second semantic vector corresponding to the same historical recommendation information and the features of the second historical behavior information. By fusing the second semantic vector with the features of the second historical behavior information, the computer device can enable each piece of historical recommendation information in the historical recommendation information sequence to correspond to the second semantic vector with the features of the second historical behavior information added.
Optionally, the computer device processes each second historical behavior information in the second historical behavior information sequence through the behavior coding network, and can obtain a historical behavior coding vector corresponding to the second historical behavior information. In the fusion process, the computer device fuses the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the historical behavior coding vector of the second piece of historical behavior information corresponding to the historical recommendation information, so as to obtain fusion characteristics. The historical behavior coding vector can reflect whether the sample user account clicks, approves, shares, collects and reviews the historical recommendation information, and can also reflect specific interaction time.
Optionally, the computer device can implement fusion of the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence and the historical behavior coding vector of the second piece of historical behavior information corresponding to the historical recommendation information through the behavior fusion network, so as to obtain the fusion feature.
In the process of fusing through the behavior fusion network, the computer device performs fusion calculation on the second semantic vector and the characteristics of the second historical behavior information, wherein the fusion calculation mode comprises addition, multiplication, fusion by using a neural network and the like. And (4) fusing the neural network, namely inputting the neural network as the features of the second semantic vector and the second historical behavior information, and outputting the neural network as fused features.
Step 312: and predicting the predicted behavior information of the exposure recommendation information through a behavior prediction network based on the first semantic vector and the fusion features corresponding to each piece of historical recommendation information in the historical recommendation information sequence.
The predicted behavior information is used to reflect which interaction behaviors may be generated by the sample user account and the exposure recommendation information. The behavior prediction network can predict which interaction behaviors possibly generated by the user account and the certain exposure recommendation information according to the characteristics of the certain exposure recommendation information recommended to the user account, the characteristics of the historical recommendation information recommended to the user account before the certain exposure recommendation information is recommended, and the interaction behaviors corresponding to the historical recommendation information.
Through the behavior prediction network, the computer equipment can predict which interaction behaviors are likely to occur between the sample user account and the exposure recommendation information based on the output of the behavior fusion network and the first semantic vector of the exposure recommendation information. Such as whether the sample user account will click, like, share, and comment on exposure recommendation information, etc.
Optionally, the computer device is capable of predicting the predicted behavior information through a neural network from the fused features and the first semantic vector based on an attention mechanism through a behavior prediction network. Alternatively, the computer device can pool (Pooling) the fused features, then perform a point multiplication with the first semantic vector, and predict the predicted behavior information through a neural network based on the output result. Alternatively, the computer device can directly connect and input the fused feature and the first semantic vector to the neural network to predict the predicted behavior information.
Step 314: training the machine learning model according to a difference between the first historical behavior information and the predicted behavior information.
The feature extraction network, the behavior prediction network, the behavior coding network and the behavior fusion network are all sub-networks of a machine learning model. The method comprises the steps of extracting features, predicting behaviors, encoding behaviors and fusing behaviors, wherein the features are extracted from a feature extraction network, the behavior prediction network, the behavior encoding network and the behavior fusion network in a cascade mode. And training a machine learning model, namely training the characteristic extraction network, the behavior prediction network, the behavior coding network and the behavior fusion network.
In training the machine learning model, the computer device can determine an error loss based on a difference between the first historical behavior information and the predicted behavior information. Training of the machine learning model can then be achieved based on back propagation through the error loss. For example, the structures of the feature extraction network, the behavior prediction network, the behavior coding network, and the behavior fusion network in the machine learning model can refer to the example in fig. 1.
Through a feature extraction network in a machine learning model, semantic vectors of reference recommendation information which is interested in a user account in the past and semantic vectors of candidate recommendation information can be extracted. By matching the semantic vector of the reference recommendation information which is interested in the user account in the past with the semantic vector of the candidate recommendation information, the recommendation information can be determined from the candidate recommendation information and recommended to the user account.
In summary, in the method provided in this embodiment, the semantic vector of the information can be extracted through the feature extraction network obtained through training, so that the information of interest can be recommended to the user based on the matching result of the semantic vector. In the process of training the feature extraction network, the used training data is data in an actual application scene, so that the accuracy of the extracted semantic vector can be improved, the accuracy when recommending information to a user is improved, and the accuracy when recommending image-text combination to the user can be improved.
In addition, in the process of training the machine learning model, the machine learning model is trained based on the processing result of the existing data without manually marking samples, so that the efficiency of training the feature extraction network and recommending information can be improved. The machine learning model for extracting the semantic vector of the information can learn the relationship between the characteristics of certain information browsed by a user, the characteristics of information browsed or interacted by the user before the certain information, the characteristics of historical behavior information of the browsed or interacted information and the interaction behavior generated by the certain information and the user, so that the machine learning model can extract the characteristics related to the behavior of the user in the information and is beneficial to improving the accuracy of subsequent information recommendation. By extracting the historical behavior encoding vector and fusing the historical behavior encoding vector and the second semantic vector, the second semantic vector can be added with the corresponding characteristics of the second historical behavior information, so that the accuracy in predicting the predicted behavior information is improved.
Fig. 4 is a flowchart illustrating an information recommendation method according to an exemplary embodiment of the present application. The method may be used in a computer device. As shown in fig. 4, the method includes:
step 402: and determining reference information to be recommended corresponding to the account of the user to be recommended.
The reference information to be recommended is information in which the account of the user to be recommended is interested, and comprises any information to be recommended which generates interactive behaviors with the account of the user to be recommended and at least one of the information to be recommended which generates interactive behaviors with the account of the user matched with the characteristics of the account of the user to be recommended.
The interactive behavior comprises at least one of clicking, praise, sharing, collecting and commenting. The user accounts matched with the characteristics of the user accounts to be recommended comprise user accounts with behavior characteristics similar to the user accounts to be recommended and user accounts with attribute characteristics similar to the user accounts to be recommended. For example, 80% of the information that was praised by the user account to be recommended in the last week is also praised by the other user account, the computer device may determine that the two user account characteristics match.
Step 404: and extracting a third semantic vector referring to the information to be recommended through a feature extraction network.
The third semantic vector can reflect a feature of the reference recommendation information. The feature extraction network is obtained by training the feature extraction network provided in the above steps 302 to 314.
Step 406: and acquiring a fourth semantic vector of a plurality of candidate information to be recommended.
The fourth semantic vector is obtained by extracting the candidate information to be recommended through a feature extraction network. Optionally, the content of the reference information to be recommended and the candidate information to be recommended is a graph-text combination, the graph-text combination is a document composed of a picture and a title, and the title is used for describing the picture.
The candidate information to be recommended is information which is stored in the computer equipment and is used for recommending the account of the user to be recommended, and the candidate information to be recommended can change along with time. Therefore, the computer device can periodically extract the features of the candidate information to be recommended through the feature extraction network, so as to obtain a fourth semantic vector. The computer equipment can extract the fourth semantic vector before extracting the third semantic vector and acquire the fourth semantic vector when information recommendation is needed.
Step 408: and determining a target fourth semantic vector matched with the third semantic vector in the fourth semantic vectors of the candidate information to be recommended, and determining the candidate information to be recommended corresponding to the target fourth semantic vector as the recommendation information.
The recommendation information is used for recommending to the account of the user to be recommended. The computer device sends the recommendation information to the account of the user to be recommended. Optionally, the computer device can determine the target fourth semantic vector by calculating a similarity (or distance) between the third semantic vector and the fourth semantic vector.
And matching the third semantic vector of the reference information to be recommended with the fourth semantic vector of the candidate information to be recommended, wherein the reference information to be recommended is information which is interested in the past of the account of the user to be recommended, and therefore the candidate information to be recommended is also information which is interested in the user. The semantic vector extracted by the feature extraction network can realize the recommendation of interesting information to the user.
In a specific example, fig. 5 is a schematic diagram of a recommended teletext combination provided in an exemplary embodiment of the application. As shown in fig. 5, when a client logged in by a user account to be recommended needs to display an information recommendation interface 501, the client sends an information recommendation request to a server. The server obtains reference image-text combination 502 interacted by the account of the user to be recommended in the past according to the information recommendation request. The content of the reference text combination 502 consists of a picture of the cat and a description of the picture. The server will then extract the semantic vector of the reference teletext combination 502 and match it with the semantic vectors of the candidate teletext combinations. Thereby determining a recommended teletext combination 503 and sending it to the client. The content of the recommended image-text combination is composed of the picture of the pet and the description of the picture, and has similar semantics. The client displays a recommendation teletext combination 503 in the information recommendation interface 501.
In summary, in the method provided by this embodiment, the third semantic vector of the reference information to be recommended and the fourth semantic vector of the candidate information to be recommended are extracted through the feature extraction network, and recommendation information can be determined in the candidate information to be recommended based on the matching result of the third semantic vector and the fourth semantic vector, so that information of interest is recommended to the user. In the process of training the feature extraction network, the used training data is data in an actual application scene, so that the accuracy of semantic vectors of the extracted information can be improved, the accuracy of recommended information is improved, and the accuracy of recommending image-text combination to a user can be improved.
It should be noted that, the order of the steps of the method provided in the embodiments of the present application may be appropriately adjusted, and the steps may also be increased or decreased according to the circumstances, and any method that can be easily conceived by those skilled in the art within the technical scope disclosed in the present application shall be covered by the protection scope of the present application, and therefore, the detailed description thereof is omitted.
In a specific example, the client is a local living client, and can recommend the user of interest in image-text combination in the aspects of food, leisure, tourism, beauty and the like.
In the model training stage, the server determines the picture-text combinations of the food, leisure, tourism, beauty and the like which are recommended to the user, the picture-text combinations browsed or interacted by the user before the recommendation of the picture-text combinations, and historical interaction behavior information generated between the user and the picture-text combinations to train the feature extraction network. And the server also updates the information periodically and continues to train the feature extraction network by using the updated information.
In the recommendation stage, when the local life client needs to display a user interface with a function of recommending the user by combining pictures and texts, or the local life client starts to run, the local life client sends a recommendation request to the server. And the server determines the image-text combination which is interesting to the user in the past (the image-text combination is interacted by the user or is interacted by the user matched with the user) in the aspects of food, leisure, tourism, beauty and the like according to the user identification in the recommendation request, and extracts the semantic vector of the image-text combination which is interesting to the user in the past by using the trained feature extraction network. And then matching the extracted semantic vector with the semantic vector of the pre-extracted image-text combination to be recommended so as to obtain the matched image-text combination in the aspects of food, leisure, tourism, beauty and the like, and then sending the matched image-text combination to a client for displaying to realize the recommendation of the interested image-text combination to the user. The above process of extracting the semantic vector of the teletext combination of interest to the user in the past may be performed in advance before processing the recommendation request.
Fig. 6 is a schematic structural diagram of a training apparatus for a feature extraction network according to an exemplary embodiment of the present application. The apparatus may be for a computer device. As shown in fig. 6, the apparatus 60 includes:
the obtaining module 601 is configured to obtain exposure recommendation information and first historical behavior information, where the exposure recommendation information is recommendation information recommended to a sample user account in a sample user account set, and the first historical behavior information is used to reflect that a first interaction behavior is generated between the sample user account and the exposure recommendation information.
The obtaining module 601 is further configured to obtain a history recommendation information sequence, where the history recommendation information sequence includes history recommendation information recommended to the sample user account before exposing the recommendation information.
The obtaining module 601 is further configured to obtain a second historical behavior information sequence according to the historical recommendation information sequence, where the second historical behavior information sequence includes second historical behavior information, and the second historical behavior information is used to reflect that a second interaction behavior is generated by the sample user account and the historical recommendation information.
The training module 602 is configured to train a feature extraction network through the exposure recommendation information, the first historical behavior information, the historical recommendation information sequence, and the second historical behavior information sequence, where the feature extraction network is configured to extract the exposure recommendation information and semantic vectors of the historical recommendation information.
In an alternative design, the feature extraction network is a sub-network in a machine learning model that further includes a behavior prediction network in cascade with the feature extraction network. A training module 602 to:
and extracting a first semantic vector of the exposure recommendation information and a second semantic vector of each history recommendation information in the history recommendation information sequence through a feature extraction network.
And fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the characteristics of the second piece of historical behavior information corresponding to the historical recommendation information to obtain fused characteristics corresponding to the historical recommendation information.
And predicting the predicted behavior information of the exposure recommendation information through a behavior prediction network based on the first semantic vector and the fusion features corresponding to each piece of historical recommendation information in the historical recommendation information sequence.
Training the machine learning model according to a difference between the first historical behavior information and the predicted behavior information.
In an alternative design, the machine learning model further includes a behavior coding network, and the behavior coding network is cascaded with the feature extraction network and the behavior prediction network. A training module 602 to:
and processing each second historical behavior information in the second historical behavior information sequence through the behavior coding network to obtain a historical behavior coding vector corresponding to the second historical behavior information.
And fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the historical behavior coding vector of the second piece of historical behavior information corresponding to the historical recommendation information to obtain fusion characteristics.
In an optional design, the machine learning model further includes a behavior fusion network, and the behavior fusion network is cascaded with the behavior prediction network, the behavior coding network and the feature extraction network. A training module 602 for
And fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the historical behavior coding vector of the second piece of historical behavior information corresponding to the historical recommendation information through a behavior fusion network to obtain fusion characteristics.
In an alternative design, the content of the exposure recommendation information and the history recommendation information is a text-text combination, the text-text combination is a document composed of pictures and titles, and the titles are used for describing the pictures.
In an alternative design, the obtaining module 601 is configured to:
and acquiring at least one of a browsing information sequence and an interaction information sequence. The browsing information sequence is composed of browsing recommendation information recommended to the sample user account before the recommendation time of the exposure recommendation information, and the interactive information sequence is composed of interactive recommendation information generating a third interactive behavior with the sample user account before the recommendation time of the exposure recommendation information.
In an optional design, the interactive recommendation information includes at least one of clicked recommendation information, praised recommendation information, shared recommendation information, collected recommendation information, and commented recommendation information.
Fig. 7 is a schematic structural diagram of an information recommendation device according to an exemplary embodiment of the present application. The apparatus may be for a computer device. As shown in fig. 7, the apparatus 70 includes:
the determining module 701 is configured to determine reference to-be-recommended information corresponding to a user account to be recommended, where the reference to-be-recommended information includes at least one of information to be recommended that generates an interactive behavior with the user account to be recommended and information to be recommended that generates an interactive behavior with the user account matched with characteristics of the user account to be recommended.
An extracting module 702, configured to extract, through the feature extraction network, a third semantic vector referring to the information to be recommended.
The obtaining module 703 is configured to obtain a fourth semantic vector of the multiple candidate information to be recommended, where the fourth semantic vector is obtained by extracting the candidate information to be recommended through a feature extraction network.
The determining module 701 is further configured to determine, in a fourth semantic vector of the multiple candidate information to be recommended, a target fourth semantic vector matched with the third semantic vector, and determine candidate information to be recommended corresponding to the target fourth semantic vector as recommendation information, where the recommendation information is used for recommending to a user account to be recommended.
In an optional design, the content of the reference information to be recommended and the candidate information to be recommended is a graph-text combination, the graph-text combination is a document composed of pictures and titles, and the titles are used for describing the pictures.
It should be noted that: the training device of the feature extraction network provided in the above embodiment is only illustrated by the division of the above functional modules, and in practical applications, the above function distribution may be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the above described functions. In addition, the training apparatus of the feature extraction network and the training method of the feature extraction network provided in the above embodiments belong to the same concept, and specific implementation processes thereof are described in detail in the method embodiments and are not described herein again.
Similarly, the information recommendation apparatus provided in the foregoing embodiment is only illustrated by dividing the functional modules, and in practical applications, the functions may be distributed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules, so as to complete all or part of the functions described above. In addition, the information recommendation device and the information recommendation method provided by the above embodiments belong to the same concept, and specific implementation processes thereof are detailed in the method embodiments and are not described herein again.
Embodiments of the present application further provide a computer device, including: the system comprises a processor and a memory, wherein at least one instruction, at least one program, a code set or an instruction set is stored in the memory, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to realize the training method or the information recommendation method of the feature extraction network provided by the method embodiments.
Optionally, the computer device is a server. Illustratively, fig. 8 is a schematic structural diagram of a computer device provided in an exemplary embodiment of the present application.
The computer apparatus 800 includes a Central Processing Unit (CPU) 801, a system Memory 804 including a Random Access Memory (RAM) 802 and a Read-Only Memory (ROM) 803, and a system bus 805 connecting the system Memory 804 and the CPU 801. The computer device 800 also includes a basic Input/Output system (I/O system) 806, which facilitates transfer of information between devices within the computer device, and a mass storage device 807 for storing an operating system 813, application programs 814, and other program modules 815.
The basic input/output system 806 includes a display 808 for displaying information and an input device 809 such as a mouse, keyboard, etc. for user input of information. Wherein the display 808 and the input device 809 are connected to the central processing unit 801 through an input output controller 810 connected to the system bus 805. The basic input/output system 806 may also include an input/output controller 810 for receiving and processing input from a number of other devices, such as a keyboard, mouse, or electronic stylus. Similarly, input-output controller 810 also provides output to a display screen, a printer, or other type of output device.
The mass storage device 807 is connected to the central processing unit 801 through a mass storage controller (not shown) connected to the system bus 805. The mass storage device 807 and its associated computer-readable storage media provide non-volatile storage for the computer device 800. That is, the mass storage device 807 may include a computer-readable storage medium (not shown) such as a hard disk or Compact Disc-Only Memory (CD-ROM) drive.
Without loss of generality, the computer-readable storage media may include computer storage media and communication media. Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable storage instructions, data structures, program modules or other data. Computer storage media includes RAM, ROM, Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), flash Memory or other solid state Memory devices, CD-ROM, Digital Versatile Disks (DVD), or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that the computer storage media is not limited to the foregoing. The system memory 804 and mass storage 807 described above may be collectively referred to as memory.
The memory stores one or more programs configured to be executed by the one or more central processing units 801, the one or more programs containing instructions for implementing the method embodiments described above, and the central processing unit 801 executing the one or more programs to implement the methods provided by the various method embodiments described above.
According to various embodiments of the present application, the computer device 800 may also operate as a remote computer device connected to a network via a network, such as the Internet. That is, the computer device 800 may be connected to the network 812 through the network interface unit 811 coupled to the system bus 805, or may be connected to other types of networks or remote computer device systems (not shown) using the network interface unit 811.
The memory also includes one or more programs, stored in the memory, that include instructions for performing the steps performed by the computer device in the methods provided by the embodiments of the present application.
The embodiment of the present application further provides a computer-readable storage medium, where at least one instruction, at least one program, a code set, or a set of instructions is stored in the computer-readable storage medium, and when the at least one instruction, the at least one program, the code set, or the set of instructions is loaded and executed by a processor of a computer device, the method for training a feature extraction network or the method for recommending information provided by the above method embodiments is implemented.
The present application also provides a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the training method or the information recommendation method of the feature extraction network provided by the above method embodiments.
It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program instructing relevant hardware, and the program may be stored in a computer readable storage medium, and the above readable storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.
The above description is only an example of the present application and should not be taken as limiting, and any modifications, equivalent switches, improvements, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims (13)

1. A method of training a feature extraction network, the method comprising:
acquiring exposure recommendation information and first historical behavior information, wherein the exposure recommendation information is recommended to a sample user account in a sample user account set, and the first historical behavior information is used for reflecting that the sample user account and the exposure recommendation information generate a first interactive behavior;
acquiring a historical recommendation information sequence, wherein the historical recommendation information sequence comprises historical recommendation information recommended to the sample user account before the exposure recommendation information;
acquiring a second historical behavior information sequence according to the historical recommendation information sequence, wherein the second historical behavior information sequence comprises second historical behavior information which is used for reflecting that the sample user account and the historical recommendation information generate a second interaction behavior;
training a feature extraction network through the exposure recommendation information, the first historical behavior information, the historical recommendation information sequence and the second historical behavior information sequence, wherein the feature extraction network is used for extracting semantic vectors of the exposure recommendation information and the historical recommendation information.
2. The method of claim 1, wherein the feature extraction network is a sub-network in a machine learning model, the machine learning model further comprising a behavior prediction network cascaded with the feature extraction network;
the training of the feature extraction network through the exposure recommendation information, the first historical behavior information, the historical recommendation information sequence, and the second historical behavior information sequence includes:
extracting a first semantic vector of the exposure recommendation information and a second semantic vector of each history recommendation information in the history recommendation information sequence through the feature extraction network;
fusing a second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the characteristics of second historical behavior information corresponding to the historical recommendation information to obtain fused characteristics corresponding to the historical recommendation information;
predicting predicted behavior information of the exposure recommendation information through the behavior prediction network based on the first semantic vector and fusion features corresponding to each piece of historical recommendation information in the historical recommendation information sequence;
training the machine learning model according to a difference between the first historical behavior information and the predicted behavior information.
3. The method of claim 2, wherein the machine learning model further comprises a behavior coding network cascaded with the feature extraction network and the behavior prediction network;
the fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the feature of the second piece of historical behavior information corresponding to the historical recommendation information to obtain the fused feature corresponding to the historical recommendation information includes:
processing each second historical behavior information in the second historical behavior information sequence through the behavior coding network to obtain a historical behavior coding vector corresponding to the second historical behavior information;
and fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the historical behavior coding vector of the second piece of historical behavior information corresponding to the historical recommendation information to obtain the fusion characteristic.
4. The method of claim 3, wherein the machine learning model further comprises a behavior fusion network cascaded with the behavior prediction network, the behavior encoding network, and the feature extraction network;
the fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the historical behavior coding vector of the second piece of historical behavior information corresponding to the historical recommendation information to obtain the fused feature includes:
and fusing the second semantic vector of each piece of historical recommendation information in the historical recommendation information sequence with the historical behavior coding vector of the second piece of historical behavior information corresponding to the historical recommendation information through the behavior fusion network to obtain the fusion characteristic.
5. The method according to any one of claims 1 to 4,
the contents of the exposure recommendation information and the history recommendation information are image-text combinations, the image-text combinations are documents formed by pictures and titles, and the titles are used for describing the pictures.
6. The method according to any one of claims 1 to 4, wherein the obtaining of the historical recommendation information sequence comprises:
acquiring at least one of a browsing information sequence and an interactive information sequence;
the browsing information sequence is composed of browsing recommendation information recommended to the sample user account before the recommendation time of the exposure recommendation information, and the interaction information sequence is composed of interaction recommendation information generating a third interaction behavior with the sample user account before the recommendation time of the exposure recommendation information.
7. The method of claim 6, wherein the interactive recommendation information comprises at least one of clicked recommendation information, praised recommendation information, shared recommendation information, favorite recommendation information, and commented recommendation information.
8. An information recommendation method applied to a computer device running a feature extraction network, the feature extraction network being trained by the method of any one of claims 1 to 7, the method comprising:
determining reference to-be-recommended information corresponding to a user account to be recommended, wherein the reference to-be-recommended information comprises at least one of information to be recommended which generates interactive behaviors with the user account to be recommended and information to be recommended which generates interactive behaviors with the user account matched with the characteristics of the user account to be recommended;
extracting a third semantic vector of the reference information to be recommended through the feature extraction network;
obtaining a fourth semantic vector of a plurality of candidate information to be recommended, wherein the fourth semantic vector is obtained by extracting the candidate information to be recommended through the feature extraction network;
and determining a target fourth semantic vector matched with the third semantic vector in fourth semantic vectors of the candidate information to be recommended, and determining the candidate information to be recommended corresponding to the target fourth semantic vector as recommendation information, wherein the recommendation information is used for recommending to the user account to be recommended.
9. The method of claim 8,
the content of the reference information to be recommended and the candidate information to be recommended is a graph-text combination, the graph-text combination is a document formed by pictures and titles, and the titles are used for describing the pictures.
10. An apparatus for training a feature extraction network, the apparatus comprising:
the system comprises an acquisition module, a processing module and a display module, wherein the acquisition module is used for acquiring exposure recommendation information and first historical behavior information, the exposure recommendation information is recommended to a sample user account in a sample user account set, and the first historical behavior information is used for reflecting that the sample user account and the exposure recommendation information generate a first interaction behavior;
the acquisition module is further configured to acquire a historical recommendation information sequence, where the historical recommendation information sequence includes historical recommendation information recommended to the sample user account before the exposure recommendation information;
the obtaining module is further configured to obtain a second historical behavior information sequence according to the historical recommendation information sequence, where the second historical behavior information sequence includes second historical behavior information, and the second historical behavior information is used to reflect that the sample user account and the historical recommendation information generate a second interaction behavior;
and the training module is used for training a feature extraction network through the exposure recommendation information, the first historical behavior information, the historical recommendation information sequence and the second historical behavior information sequence, wherein the feature extraction network is used for extracting the exposure recommendation information and semantic vectors of the historical recommendation information.
11. An information recommendation apparatus, wherein the apparatus is operated with a feature extraction network, the feature extraction network being a network trained by the apparatus of claim 10, the apparatus comprising:
the system comprises a determining module, a recommending module and a recommending module, wherein the determining module is used for determining reference to-be-recommended information corresponding to a user account to be recommended, and the reference to-be-recommended information comprises at least one of information to be recommended which generates an interactive behavior with the user account to be recommended and is matched with the characteristics of the user account to be recommended;
the extraction module is used for extracting a third semantic vector of the reference information to be recommended through the feature extraction network;
the acquisition module is used for acquiring a fourth semantic vector of a plurality of candidate information to be recommended, and the fourth semantic vector is obtained by extracting the candidate information to be recommended through the feature extraction network;
the determining module is further configured to determine, in a fourth semantic vector of the multiple candidate information to be recommended, a target fourth semantic vector matched with the third semantic vector, and determine candidate information to be recommended corresponding to the target fourth semantic vector as recommendation information, where the recommendation information is used for recommending to the user account to be recommended.
12. A computer device comprising a processor and a memory, the memory having stored therein at least one instruction, at least one program, a set of codes, or a set of instructions, the at least one instruction, the at least one program, the set of codes, or the set of instructions being loaded and executed by the processor to implement the method of training a feature extraction network according to any one of claims 1 to 7, or the method of information recommendation according to claim 8 or 9.
13. A computer-readable storage medium, wherein at least one instruction, at least one program, a set of codes, or a set of instructions is stored in the computer-readable storage medium, and the at least one instruction, the at least one program, the set of codes, or the set of instructions is loaded and executed by a processor to implement the method for training the feature extraction network according to any one of claims 1 to 7, or the method for recommending information according to claim 8 or 9.
CN202110902980.4A 2021-08-06 2021-08-06 Training method of feature extraction network, information recommendation method, device and equipment Active CN113672820B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110902980.4A CN113672820B (en) 2021-08-06 2021-08-06 Training method of feature extraction network, information recommendation method, device and equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110902980.4A CN113672820B (en) 2021-08-06 2021-08-06 Training method of feature extraction network, information recommendation method, device and equipment

Publications (2)

Publication Number Publication Date
CN113672820A true CN113672820A (en) 2021-11-19
CN113672820B CN113672820B (en) 2022-09-16

Family

ID=78541756

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110902980.4A Active CN113672820B (en) 2021-08-06 2021-08-06 Training method of feature extraction network, information recommendation method, device and equipment

Country Status (1)

Country Link
CN (1) CN113672820B (en)

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150309714A1 (en) * 2014-04-23 2015-10-29 Oracle International Corporation Navigating interactive visualizations with collaborative filtering
CN107577737A (en) * 2017-08-25 2018-01-12 北京百度网讯科技有限公司 Method and apparatus for pushed information
CN110276446A (en) * 2019-06-26 2019-09-24 北京百度网讯科技有限公司 The method and apparatus of model training and selection recommendation information
CN111241381A (en) * 2018-11-28 2020-06-05 北京奇虎科技有限公司 Information recommendation method and device, electronic equipment and computer-readable storage medium
CN111310056A (en) * 2020-03-11 2020-06-19 腾讯科技（深圳）有限公司 Information recommendation method, device, equipment and storage medium based on artificial intelligence
CN111680217A (en) * 2020-05-27 2020-09-18 腾讯科技（深圳）有限公司 Content recommendation method, device, equipment and storage medium
CN111831924A (en) * 2020-07-16 2020-10-27 腾讯科技（北京）有限公司 Content recommendation method, device, equipment and readable storage medium
US20210133558A1 (en) * 2019-10-31 2021-05-06 International Business Machines Corporation Deep-learning model creation recommendations

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150309714A1 (en) * 2014-04-23 2015-10-29 Oracle International Corporation Navigating interactive visualizations with collaborative filtering
CN107577737A (en) * 2017-08-25 2018-01-12 北京百度网讯科技有限公司 Method and apparatus for pushed information
CN111241381A (en) * 2018-11-28 2020-06-05 北京奇虎科技有限公司 Information recommendation method and device, electronic equipment and computer-readable storage medium
CN110276446A (en) * 2019-06-26 2019-09-24 北京百度网讯科技有限公司 The method and apparatus of model training and selection recommendation information
US20210133558A1 (en) * 2019-10-31 2021-05-06 International Business Machines Corporation Deep-learning model creation recommendations
CN111310056A (en) * 2020-03-11 2020-06-19 腾讯科技（深圳）有限公司 Information recommendation method, device, equipment and storage medium based on artificial intelligence
CN111680217A (en) * 2020-05-27 2020-09-18 腾讯科技（深圳）有限公司 Content recommendation method, device, equipment and storage medium
CN111831924A (en) * 2020-07-16 2020-10-27 腾讯科技（北京）有限公司 Content recommendation method, device, equipment and readable storage medium

Also Published As

Publication number Publication date
CN113672820B (en) 2022-09-16

Similar Documents

Publication Publication Date Title
CN111444428B (en) Information recommendation method and device based on artificial intelligence, electronic equipment and storage medium
CN108921221B (en) User feature generation method, device, equipment and storage medium
CN108304439B (en) Semantic model optimization method and device, intelligent device and storage medium
CN108776676B (en) Information recommendation method and device, computer readable medium and electronic device
CN111680219A (en) Content recommendation method, device, equipment and readable storage medium
CN111090756B (en) Artificial intelligence-based multi-target recommendation model training method and device
CN110234018B (en) Multimedia content description generation method, training method, device, equipment and medium
CN109271509B (en) Live broadcast room topic generation method and device, computer equipment and storage medium
CN110413867B (en) Method and system for content recommendation
CN108062377A (en) The foundation of label picture collection, definite method, apparatus, equipment and the medium of label
CN111046275A (en) User label determining method and device based on artificial intelligence and storage medium
CN112989212B (en) Media content recommendation method, device and equipment and computer storage medium
CN113010702A (en) Interactive processing method and device for multimedia information, electronic equipment and storage medium
CN111831924A (en) Content recommendation method, device, equipment and readable storage medium
CN110737824B (en) Content query method and device
CN111460290A (en) Information recommendation method, device, equipment and storage medium
CN115329131A (en) Material label recommendation method and device, electronic equipment and storage medium
CN113947131A (en) Model training method, device, equipment and storage medium
CN113704509A (en) Multimedia recommendation method and device, electronic equipment and storage medium
CN114817692A (en) Method, device and equipment for determining recommended object and computer storage medium
CN113032676A (en) Recommendation method and system based on micro-feedback
CN116956183A (en) Multimedia resource recommendation method, model training method, device and storage medium
CN113672820B (en) Training method of feature extraction network, information recommendation method, device and equipment
CN114357242A (en) Training evaluation method and device based on recall model, equipment and storage medium
CN111310016B (en) Label mining method, device, server and 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