Recommended method, device, computing device and storage medium

Abstract

The embodiments of the present disclosure provide a recommendation method, apparatus, computing device and storage medium. The method includes: obtaining the attribute characteristics of a first user; inputting the attribute characteristics of the first user into a cross-reconstruction network, cross-processing the attribute characteristics of the first user using the cross-reconstruction network to obtain the behavior characteristics of the first user; obtaining the score of the media information according to the attribute characteristics, behavior characteristics and characteristics of the media information of the first user; recommending the first user according to the score of the media information, cross-processing the attribute characteristics of the first user using the cross-reconstruction network so that the encoding characteristics of the attribute characteristics have the same predictive ability as the encoding characteristics of the behavior characteristics, thereby being able to predict the behavior characteristics of the first user, constructing a virtual behavior characteristic, and thus achieving personalized recommendation for the first user.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of information processing technology. More specifically, the embodiments of the present disclosure relate to a recommendation method, apparatus, computing device and storage medium. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the disclosure that are recited in the claims. No description herein is admitted to be prior art by inclusion in this section.

[0003] With the development of Internet technology, various Internet platforms have a large amount of information resources. For example, music platforms can provide users with a large amount of music resources, e-commerce platforms have a large amount of commodity resources, etc. Platforms usually analyze users' interests and preferences by collecting their behavior data, and recommend resources that users may be interested in based on this.

[0004] In the absence of user behavior, platforms usually rely on a large number of prior rules to classify users and make recommendations based on the user classification results. This approach introduces a lot of manual work and causes popular items to have higher weights among users of various characteristics, resulting in a lack of personalization in the overall recommendation results. Summary of the invention

[0005] In view of this, the embodiments of the present disclosure at least provide a recommendation method, apparatus, computing device and storage medium, which can overcome tedious manual work and achieve personalized recommendations for new users who lack behavioral data.

[0006] In a first aspect of an embodiment of the present disclosure, a recommendation method is provided, comprising: obtaining attribute characteristics of a first user, wherein the first user is a user without behavioral characteristics; inputting the attribute characteristics of the first user into a cross-reconstruction network, cross-processing the attribute characteristics of the first user using the cross-reconstruction network to obtain the behavioral characteristics of the first user; obtaining a score of the media information based on the attribute characteristics, behavioral characteristics and characteristics of the media information of the first user; and recommending the first user based on the score of the media information.

[0007] In one embodiment of the present disclosure, the cross-reconstruction network includes a first autoencoder and a second autoencoder, and the cross-reconstruction network is used to cross-process the attribute characteristics of the first user to obtain the behavior characteristics of the first user, including: inputting the attribute characteristics of the first user into the first autoencoder, encoding the attribute characteristics using the first autoencoder to obtain target encoded characteristics; and decoding the target encoded characteristics using the second autoencoder to obtain the behavior characteristics of the first user.

[0008] In one embodiment of the present disclosure, obtaining the score of the media information based on the attribute characteristics, behavioral characteristics and characteristics of the media information of the first user includes: inputting the attribute characteristics, behavioral characteristics and characteristics of the media information of the first user into an index prediction network to obtain a predicted value of a target index of the media information, wherein the index prediction network includes a fully connected layer, and the target index includes at least one of the following: click-through rate, conversion rate, and click-through conversion rate.

[0009] In one embodiment of the present disclosure, the recommending the first user based on the score of the media information includes: sorting each media information according to the predicted value of the target index of each media information in the set of media information to be recommended; and recommending the media information based on the sorting result.

[0010] In one embodiment of the present disclosure, the method further includes: obtaining attribute characteristics and behavior characteristics of a second user, wherein the second user is a user with behavior characteristics; and training the cross-reconstruction network using the attribute characteristics and behavior characteristics of the second user.

[0011] In one embodiment of the present disclosure, the cross reconstruction network includes a first encoder, a second encoder, a first decoder and a second decoder, wherein the first encoder is used to encode an input attribute feature, the second encoder is used to encode an input behavior feature, the first decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, and the second decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, and the training loss of the cross reconstruction network includes: a first loss, which is used to indicate the difference between the result obtained by the first decoder decoding the first encoding feature and the attribute feature of the second user, and the difference between the result obtained by the second decoder decoding the first encoding feature and the behavior feature of the second user; wherein the first encoding feature is obtained by the first encoder encoding the attribute feature of the second user; a second loss, which is used to indicate the difference between the result obtained by the first decoder decoding the second encoding feature and the attribute feature of the second user, and the difference between the result obtained by the second decoder decoding the second encoding feature and the behavior feature of the second user; wherein the second encoding feature is obtained by the second encoder encoding the behavior feature of the second user.

[0012] In one embodiment of the present disclosure, the training loss of the cross reconstruction network also includes: a third loss, used to indicate the average difference between a first mapping result and a second mapping result, wherein the first mapping result is obtained by mapping the first encoding feature to a reproducing kernel Hilbert space, and the second mapping result is obtained by mapping the second encoding feature to a reproducing kernel Hilbert space.

[0013] In one embodiment of the present disclosure, the cross-reconstruction network also includes: a first attention mechanism network, used to map the embedding vector contained in the attribute features of the target user into a hidden attribute vector, and replace the attribute features with the hidden attribute vector, wherein the target user includes a first user and a second user; a second attention mechanism network, used to map the embedding vector contained in the behavior features of the second user into a hidden behavior vector, and replace the behavior features with the hidden behavior vector.

[0014] In one embodiment of the present disclosure, the method further includes: jointly training the cross reconstruction network and the exponential prediction network.

[0015] In a second aspect of the disclosed embodiments, a recommendation device is provided, comprising: an acquisition unit, configured to acquire attribute characteristics of a first user, wherein the first user is a user without behavioral characteristics; a cross-reconstruction unit, configured to input the attribute characteristics of the first user into a cross-reconstruction network, and use the cross-reconstruction network to cross-process the attribute characteristics of the first user to obtain the behavioral characteristics of the first user; a prediction unit, configured to obtain a score of the media information based on the attribute characteristics, behavioral characteristics and characteristics of the media information of the first user; and a recommendation unit, configured to recommend the first user based on the score of the media information.

[0016] In one embodiment of the present disclosure, the cross-reconstruction network includes a first autoencoder and a second autoencoder, and the cross-reconstruction unit is specifically used to: input the attribute characteristics of the first user into the first autoencoder, encode the attribute characteristics using the autoencoder to obtain target encoded characteristics; and decode the target encoded characteristics using the second autodecoder to obtain the behavioral characteristics of the first user.

[0017] In one embodiment of the present disclosure, the prediction unit is specifically used to: input the attribute characteristics, behavioral characteristics and characteristics of the first user and the media information into an index prediction network to obtain a predicted value of a target index of the media information, wherein the index prediction network includes a fully connected layer, and the target index includes at least one of the following: click-through rate, conversion rate, and click-to-conversion rate.

[0018] In one embodiment of the present disclosure, the recommendation unit is specifically configured to: sort each piece of media information in the to-be-recommended media information set according to a predicted value of a target index of each piece of media information; and recommend the media information according to the sorting result.

[0019] In one embodiment of the present disclosure, the device also includes a first training unit, which is used to: obtain attribute characteristics and behavior characteristics of a second user, wherein the second user is a user with behavioral characteristics; and train the cross-reconstruction network using the attribute characteristics and behavior characteristics of the second user.

[0020] In one embodiment of the present disclosure, the cross-reconstruction network includes a first encoder, a second encoder, a first decoder and a second decoder, wherein the first encoder is used to encode an input attribute feature, the second encoder is used to encode an input behavior feature, the first decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, and the second decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, and the training loss of the cross-reconstruction network includes: a first loss, which is used to indicate the difference between the result obtained by the first decoder decoding the first encoding feature and the attribute feature of the second user, and the difference between the result obtained by the second decoder decoding the first encoding feature and the behavior feature of the second user; wherein the first encoding feature is obtained by the first encoder encoding the attribute feature of the second user; a second loss, which is used to indicate the difference between the result obtained by the first decoder decoding the second encoding feature and the attribute feature of the second user, and the difference between the result obtained by the second decoder decoding the second encoding feature and the behavior feature of the second user; wherein the second encoding feature is obtained by the second encoder encoding the behavior feature of the user.

[0021] In one embodiment of the present disclosure, the training loss of the cross reconstruction network also includes: a third loss, used to indicate the average difference between a first mapping result and a second mapping result, wherein the first mapping result is obtained by mapping the first encoding feature to a reproducing kernel Hilbert space, and the second mapping result is obtained by mapping the second encoding feature to a reproducing kernel Hilbert space.

[0022] In one embodiment of the present disclosure, the cross-reconstruction network also includes: a first attention mechanism network, used to map the embedding vector contained in the attribute features of the target user into a hidden attribute vector, and replace the attribute features with the hidden attribute vector, wherein the target user includes a first user and a second user; a second attention mechanism network, used to map the embedding vector contained in the behavior features of the second user into a hidden behavior vector, and replace the behavior features with the hidden behavior vector.

[0023] In one embodiment of the present disclosure, the apparatus further includes a joint training unit for jointly training the cross reconstruction network and the exponential prediction network.

[0024] In a third aspect of the embodiments of the present disclosure, a computing device is provided, including: a processor; and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the method described in any embodiment of the present disclosure is implemented.

[0025] In a fourth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, comprising: a computer program, wherein when the computer program is executed by a processor, the method described in any embodiment of the present disclosure is implemented.

[0026] According to the recommendation method, apparatus, computing device and storage medium of the embodiment of the present disclosure, for a first user without behavioral characteristics, the attribute characteristics of the first user are cross-processed using a cross-reconstruction network to obtain the behavioral characteristics of the first user, and the score of the media information is obtained based on the attribute characteristics, behavioral characteristics and characteristics of the media information of the first user to recommend the first user. By processing the attribute characteristics of the first user using a cross-reconstruction network, the encoding characteristics of the attribute characteristics and the encoding characteristics of the behavioral characteristics have the same prediction ability, thereby enabling the prediction of the behavioral characteristics of the first user and constructing a virtual behavioral characteristic, thereby achieving personalized recommendation for the first user. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0028] Figure 1 Schematically shows a flow chart of a recommended method according to an embodiment of the present disclosure;

[0029] Figure 2A The system architecture diagram schematically shows the application of the recommended method according to the embodiment of the present disclosure;

[0030] Figure 2B The following schematically shows the structure of an index prediction network according to an embodiment of the present disclosure;

[0031] Figure 3 The structure of a recommendation device according to an embodiment of the present disclosure is schematically shown;

[0032] Figure 4 Schematically illustrates a computer-readable storage medium according to an embodiment of the present disclosure;

[0033] Figure 5 The schematic diagram shows a structure of a computing device according to an embodiment of the present disclosure.

[0034] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0035] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0036] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0037] According to the embodiments of the present disclosure, a recommended method, apparatus, computing device and storage medium are proposed. In this document, it should be understood that any number of elements in the drawings is for illustration and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0038] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure.

[0039] When a new user enters a platform, network or APP, since there is no historical behavior data about the user in the recommendation system, personalized recommendations for new users (users with no behavioral characteristics or behavior information) are often referred to as the cold start problem.

[0040] In the related art, a common cold start solution is to classify the registration information of new users and recommend the users' favorite products, music, videos, etc. in the category according to the category to which the new users belong. The registration information of new users usually includes the new users' age, gender, place of residence, occupation, education, etc. However, this solution relies on a large number of prior rules to classify new users, which introduces a lot of manual work, and the actual effect also depends on the accuracy of user classification; and this method will cause popular items to have a higher weight among users of various characteristics, resulting in a lack of personalization in the overall recommendation results, which also aggravates the Matthew effect of the recommendation system.

[0041] Therefore, the recommendation scheme proposed in the embodiment of the present disclosure is intended to overcome tedious manual work and achieve personalized recommendations for new users who lack behavioral data.

[0042] Figure 1 The flow of a recommendation method according to an embodiment of the present disclosure is schematically shown. The recommendation method can be executed by an electronic device such as a terminal device or a server. The terminal device can be a fixed terminal or a mobile terminal, such as a mobile phone, a tablet computer, a game console, a desktop computer, an advertising machine, an all-in-one machine, a vehicle terminal, etc. The server includes a local server or a cloud server, etc. The method can also be implemented by a processor calling a computer-readable instruction stored in a memory. Figure 1 As shown, the method may include the following processing:

[0043] In step 101, attribute characteristics of a first user are obtained.

[0044] The first user is a user without behavioral characteristics. For example, it can be a new user who logs into the platform, network or APP for the first time, or a user who logs into the platform, network or APP as a visitor without logging in, and the recommendation system does not store historical behavior information about the user.

[0045] The attribute characteristics of the first user may be obtained by extracting the characteristics of the attribute information of the first user, and the attribute information may include multiple characteristics related to the first user, including age, gender, place of residence, occupation, education, etc. The attribute information may be obtained through the registration information of the first user, or may be obtained according to the information corresponding to the terminal used by the first user. Those skilled in the art should understand that the attribute characteristics may also include characteristics obtained by other means, and the present disclosure does not limit this.

[0046] In step 102, the attribute characteristics of the first user are input into a cross reconstruction network, and the attribute characteristics of the first user are cross-processed using the cross reconstruction network to obtain the behavior characteristics of the first user.

[0047] In the disclosed embodiment, the attribute information of the first user may be cross-processed by a trained cross-reconstruction network, and the behavior characteristics of the first user may be predicted based on the attribute characteristics. The predicted behavior characteristics may also be referred to as virtual behavior characteristics.

[0048] Among them, cross processing refers to using one of the dual autoencoders to encode the attribute feature rows, and using the other autoencoder to decode the encoded features of the attribute features. During the training process, the dual autoencoders are used to construct and learn the user attribute space and the user behavior space, and cross-modal reconstruction is performed in the public hidden space of the autoencoder, and the modal transfer relationship between the user behavior and user attributes of the old users is established. Since user attributes are homogeneous, the behavior of new users can be reconstructed based on the modal transfer relationship between the user behavior and user attributes of the old users.

[0049] In step 103, a score of the media information is obtained according to the attribute characteristics and behavior characteristics of the first user and the characteristics of the media information.

[0050] When the attribute characteristics of the first user are known and the behavioral characteristics of the first user are predicted, the first user's interest level in the media information can be determined based on the attribute information, behavioral information, and characteristics of the media information, wherein the interest level can be represented by the media information score.

[0051] In step 104, the first user is recommended according to the score of the media information.

[0052] According to the scores of the media information, at least one piece of media information that the first user is interested in may be determined from among all the media information, so as to be recommended to the first user.

[0053] In an embodiment of the present disclosure, for a first user who does not have behavioral characteristics, a cross-reconstruction network is used to cross-process the attribute characteristics of the first user to obtain the behavioral characteristics of the first user, and a score of the media information is obtained based on the attribute characteristics, behavioral characteristics and characteristics of the media information of the first user to recommend the first user. By using a cross-reconstruction network to process the attribute characteristics of the first user, the encoding characteristics of the attribute characteristics and the encoding characteristics of the behavioral characteristics have the same prediction ability, thereby enabling the prediction of the behavioral characteristics of the first user and constructing a virtual behavioral characteristic, thereby achieving personalized recommendation for the first user.

[0054] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0055] Figure 2A The system architecture diagram of the recommended method application according to the embodiment of the present disclosure is shown. Figure 2A The recommendation system shown can recommend media information to new users who enter the recommendation system for the first time, that is, first users who do not have behavior information or behavior characteristics. The recommendation system proposed in the embodiment of the present disclosure can be called a cold start system.

[0056] In the cold start scenario, the dataset of the first user can be expressed as {attribute information, media information, other information}, where the other information is the information about the first user that can be obtained, such as context, etc. Before executing the recommendation method, the dataset of the first user can be preprocessed, for example, by using a feature extraction network to extract features from the attribute information of the first user to obtain the features of the attribute information of the first user, that is, the attribute features; at the same time, the features of the media information and the features of other information are also obtained. The feature set of the first user can be expressed as x o ={α o , t o , c o}, where o represents the first user (new user), α o is the attribute feature of the first user, t o is the feature of the media information, c o A feature of the context of the first user.

[0057] like Figure 2A As shown, the recommendation system proposed in the embodiment of the present disclosure mainly includes a cross-reconstruction network 20, an index prediction network 21, and a recommendation network 22. The cross-reconstruction network 20 is used to select the first user according to the attribute feature a o , predict the behavioral feature ν of the first user o ; Index prediction network 21 is used to predict the attribute characteristics of the first user according to α o 、Behavioral characteristics o , characteristics of media information o etc., predicting the score of the media information; and the recommendation network 22 is used to make recommendations to the first user based on the score of the media information.

[0058] The following is the cross-reconstruction network 20 based on the attribute feature a of the first user: o Predictive behavioral traits o The method, that is, the application process of the cross reconstruction network is described.

[0059] like Figure 2AAs shown, the cross reconstruction network 20 includes a first autoencoder 210 and a second autoencoder 220. The first autoencoder 210 and the second autoencoder 220 can be used to cross-process the attribute features of the first user by the following method:

[0060] First, the attribute feature α of the first user o Input to the first self-encoder 210, the attribute characteristic α o It is obtained by extracting features from the attribute information of the first user.

[0061] Using the first autoencoder 210 to identify the attribute feature α o Encode to obtain the attribute feature α o The encoding feature, that is, the target encoding feature p o The first self-encoder 210 may include a first encoder 211 and a first decoder 212, wherein the attribute feature a of the first user o Input to the first encoder 211, the first encoder 211 encodes the attribute feature to obtain the target encoding feature p o .

[0062] Next, the target encoding feature p is encoded using the second autoencoder 220 o Decode to obtain the predicted behavior characteristics of the first user The second autoencoder 220 includes a second encoder 221 and a second decoder 222, and the attribute feature α of the first user o The coding feature, that is, the target coding feature p o Input to the second decoder 222, the second decoder 222 decodes the target encoding feature p o Decode to obtain the attribute feature a of the first user o Predicted behavioral characteristics In the disclosed embodiment, the predicted behavior characteristics may also be referred to as virtual behavior characteristics.

[0063] After predicting the behavior characteristics of the first user The attribute feature α of the first user o , behavioral characteristics and the characteristics of media information o The data is input to the index prediction network 21 to obtain the predicted value of the target index of the media information.

[0064] The index prediction network 21 may adopt a method of combining a deep neural network with a fully connected layer to learn cross-information of the first user's attribute characteristics, behavior characteristics, and characteristics of media information.

[0065] Figure 2B The schematic diagram shows the structure of the index prediction network according to the embodiment of the present disclosure. Figure 2B As shown, the index prediction network 21 may include a cascade network 201, fully connected layers 202, 203. The cascade network 201 cascades the input behavior feature v, attribute feature α, and feature t of the media information. In the process of predicting the target index of a new user, the input is the virtual behavior feature of the new user predicted by the cross reconstruction network. And the attribute feature α o , characteristics of media information o The fully connected layers 202 and 203 are used to perform fully connected processing on the cascaded features and output the predicted value of the target index. Those skilled in the art should understand that the number of the above fully connected layers is only an example and can be set as needed, and the embodiments of the present disclosure are not limited to this.

[0066] Other network structures related to the factorization machine and network results related to the user sequence may also be used, including but not limited to FM, DeepFM, XDeepFM, DeepCross, DIN, DIEN, DMR, etc. The structure of the exponential prediction network can be replaced according to needs, is not limited to the original network structure, and supports flexible migration.

[0067] In the embodiment of the present disclosure, the target index predicted by the index prediction network includes at least one of the following: click-through rate, conversion rate, click-through conversion rate, etc. Those skilled in the art should understand that the target index may also include other indexes that can express the first user's interest in the media information, and the embodiment of the present disclosure does not limit this.

[0068] In one example, the feature c of the context of the first user may also be o With attribute feature a o , behavioral characteristics and the characteristics of media information o The context is inputted into the index prediction network together, and the context includes, for example, the time and location of the user's visit to the recommendation system, etc. By adding the context as the input of the index prediction network, the prediction effect of the target index of the media information can be improved.

[0069] The above method can be used to obtain the predicted value of the target index of each media information in the set of media information to be recommended. Next, the recommendation network 22 can be used to sort each media information according to the predicted value of the target index of each media information, and the target media information can be sorted according to the sorting result. The media information is recommended. For example, one or more media information ranked top may be recommended to the first user as target media information.

[0070] In the disclosed embodiment, a cross-reconstruction network is used to predict virtual behavior characteristics of the new user based on the attribute characteristics of the new user, and the predicted virtual behavior characteristics, attribute characteristics, characteristics of the media information, or characteristics of other information are input into an index prediction network to obtain a predicted value of a set index of the media information, thereby determining the new user's interest in each media information in a recommended media information set. The recommendation network is then used to sort the information based on the predicted values ​​of the set index, and recommendations are made based on the sorting results. One or more media information that the new user is most likely to be interested in can be recommended, thereby solving the cold start problem of the recommendation system and achieving better recommendation effects.

[0071] The training method of the cross reconstruction network 20 is described below.

[0072] In the training method proposed in the embodiment of the present disclosure, the data of the second user (old user) with behavioral characteristics or behavioral information in the recommendation system is used to train the cross reconstruction network 20. That is, by learning the attribute-behavior links of old users from multiple modalities, the potential behavior and interests of new users are inferred, and the cold start problem of the recommendation system is solved.

[0073] The data set of the second user can be expressed as {attribute information, behavior information, media information, other information}, where the other information is the information about the second user that can be obtained, such as context, etc. Similar to the application process, the data set of the second user can be preprocessed before training, for example, using a feature extraction network to extract features from the attribute information and behavior information of the first user to obtain the attribute features and behavior features of the second user; at the same time, the features of the media information and the features of other information are also obtained. The feature set of the second user can be expressed as x s ={α s , v s , t s , c s}, where s represents the second user (old user), α s is the attribute feature of the second user, v s is the behavioral characteristic of the second user, t s is the feature of the media information, c s The feature of the context of the second user.

[0074] First, the attribute feature α of the second user sInput to the first encoder 211 in the first self-encoder 210, the first encoder 211 performs attribute feature α s Perform feature encoding to obtain attribute feature α s The first encoding feature p s At the same time, the behavioral feature v of the second user s The second encoder 221 in the second self-encoder 220 is inputted, and the second encoder 221 processes the behavior feature v s Perform feature encoding to obtain behavioral features v s The second encoding feature q s .

[0075] Next, the first encoding feature p s The first decoder 212 in the first autoencoder 210 and the second decoder 222 in the second autoencoder 220 are input simultaneously; and the second encoding feature q s The first decoder 212 in the first autoencoder 210 and the second decoder 222 in the second autoencoder 220 are input simultaneously, and the first decoder 212 encodes the first feature code p s Decode and encode the second feature q s Decode; at the same time, the second decoder 222 encodes the second feature q s Decode and encode the first pair of features p s Through the above method, the attribute feature α can be decoded. s and behavioral characteristics s Cross-modal reconstruction aligns user attributes and user behaviors by constructing cross-modal reconstruction, so that attribute-based behavior generation mapping functions can be generated more effectively to achieve behavior migration from old users to new users.

[0076] In some embodiments, the loss function of the cross reconstruction network includes at least two parts, one of which can be regarded as a first loss in the user attribute space, used to indicate the first decoder 212 to the first encoding feature p s The decoded result is consistent with the attribute feature a of the second user. s The difference between the first encoded feature p and the second decoder s The decoded result and the behavioral feature v of the second user s The difference between.

[0077] The first loss can be expressed by formula (1):

[0078]

[0079] Where D1 represents the first decoder, D2 represents the second decoder, and ps Represents the first encoding feature.

[0080] The cross reconstruction network is trained using a loss function including the first loss, so that the first encoding feature p s The attribute characteristics of the second user are reconstructed through D1 and D2 respectively and behavioral characteristics

[0081] The other part at least included in the loss function can be regarded as a second loss in the user behavior space, which is used to indicate the first decoder 212 to the second encoding feature q s The decoding result and the behavioral feature ν of the second user s The difference between the first and second decoders for the second encoded feature q s The decoded result and the behavioral feature v of the second user s The difference between.

[0082] The second loss can be expressed by formula (2):

[0083]

[0084] Where D1 represents the first decoder, D2 represents the second decoder, and q s Represents the second encoding feature.

[0085] The cross reconstruction network is trained using a loss function including the second loss, so that the second encoding feature q s The attribute characteristics of the second user are reconstructed through D1 and D2 respectively and behavioral characteristics

[0086] In the embodiment of the present disclosure, the cross-modal reconstruction network is trained by the cross-modal reconstruction loss of user attributes and behaviors, that is, the loss function of the first loss in the user attribute space and the second loss in the user behavior space, so that the first encoding feature p s and the second encoding feature q s Has the same predictive ability and can reconstruct the user's attribute features α s and ν s , so that behavioral characteristics can be predicted based on user attribute characteristics, and behavioral migration from old users to new users can be achieved.

[0087] In some embodiments, the training loss of the cross reconstruction network also includes: a third loss, used to indicate the average difference between a first mapping result and a second mapping result, wherein the first mapping result is obtained by mapping the first encoding feature to a reproducing kernel Hilbert space, and the second mapping result is obtained by mapping the second encoding feature to a reproducing kernel Hilbert space.

[0088] The third loss can be expressed by formula (3):

[0089]

[0090] in, represents the reproducing kernel Hilbert space, D represents the second decoder, φ represents the mapping function, k(.,.) represents the Hilbert space The reproducing kernel of , n represents the number of encoded features.

[0091] The third loss is used to make the first encoding feature p s and the second encoding feature q s Sharing the same hidden space and assisting the convergence of the training function of the cross reconstruction network, that is, the third loss can be used to normalize the generation ability of the second decoder for ν.

[0092] In the embodiment of the present disclosure, the training loss function of the cross reconstruction network may be:

[0093]

[0094] in, For the first loss, For the third loss, is the third loss. Through this training loss function, the effect of predicting the behavior characteristics based on the attribute characteristics of new users can be improved.

[0095] For the exponential prediction network 21, samples for training can be constructed in the following manner.

[0096] First, an identifier f is constructed to indicate a first user (new user) and a second user (old user), wherein when f=1, it indicates the first user; and when f=0, it indicates the second user.

[0097] Next, construct the input sample × according to the identifier f:

[0098]

[0099] When the flag f=1, the first user (new user) features include attribute features α o , characteristics of media information o , context features co , New user virtual behavior characteristics predicted by cross-reconstruction network When the flag f=0, the features of the second user (old user) include attribute features α s , characteristics of media information s , context features c s 、Behavioral characteristics s By labeling the true value of the target index for the feature with the identifier f=0, samples for training the index prediction network can be obtained.

[0100] In some embodiments, Figure 2A As shown, the cross-reconstruction network also includes a first attention mechanism network 231 and a second attention mechanism network 232. The first attention mechanism network 231 is used to map the embedding vector contained in the attribute feature of the second user into a hidden attribute vector And using the hidden attribute vector Replace the attribute feature and input it into the first autoencoder 210 in the cross reconstruction network 20, so that the cross reconstruction network 20 can be used according to the hidden attribute vector of the second user to predict the behavior characteristics of the second user; the second attention mechanism network 232 is used to map the embedded vector contained in the behavior characteristics of the second user into a hidden behavior vector And using the hidden behavior vector Replace the behavior feature input to the second autoencoder 220 in the cross reconstruction network 20, so that the cross reconstruction network 20 is based on the hidden behavior vector of the second user To predict the attribute characteristics of the second user.

[0101] In one example, the hidden attribute vector The reconstruction function is shown in formula (6):

[0102]

[0103] Among them, n α is the number of attribute features of the second user, α si is the embedding vector of the i-th second user, α i is the feature weight, which can be defined according to formulas (7) and (8):

[0104]

[0105] e i =z T tanh(wα si +b) (8)

[0106] in, as well as is the training parameter, d r To set the hidden vector size.

[0107] In the application process, the embedding vector contained in the attribute feature of the first user can be mapped to a hidden attribute vector And using the hidden attribute vector Replace the attribute feature and input it into the first autoencoder 210 in the cross reconstruction network 20, so that the cross reconstruction network 20 can be used according to the hidden attribute vector of the second user To predict the behavioral characteristics of the first user.

[0108] For the cross-reconstruction network and the exponential prediction network in the recommendation system in the embodiment of the present disclosure, they can be trained separately so that the cross-reconstruction network does not participate in the learning process of the exponential prediction network; they can also participate in the learning process of the exponential prediction network, and the cross-reconstruction network and the exponential prediction network are jointly trained, and a method of simultaneous segmentation or alternating optimization can be adopted.

[0109] The recommendation method proposed in the embodiment of the present disclosure can be applied to a variety of recommendation scenarios, such as homepage music recommendations or live broadcast recommendations on music platforms, product recommendations on e-commerce websites, video recommendations on video platforms or short video apps, picture and text recommendations on information flow platforms, etc.

[0110] In order to implement the recommendation method of any embodiment of the present disclosure, the embodiment of the present disclosure also provides a recommendation device. Figure 3 The structure diagram of the recommendation device according to an embodiment of the present disclosure is schematically shown. For example, the device can be applied to a computing device. In the following description, the functions of each module of the device will be briefly described, and the detailed processing can be combined with the description of the recommendation method of any embodiment of the present disclosure.

[0111] like Figure 3 As shown, the device may include: an acquisition unit 301, used to acquire attribute characteristics of a first user, wherein the first user is a user without behavioral characteristics; a cross-reconstruction unit 302, used to input the attribute characteristics of the first user into a cross-reconstruction network, and use the cross-reconstruction network to cross-process the attribute characteristics of the first user to obtain the behavioral characteristics of the first user; a prediction unit 303, used to obtain a score of the media information based on the attribute characteristics, behavioral characteristics and characteristics of the media information of the first user; and a recommendation unit 304, used to recommend the first user based on the score of the media information.

[0112] In one embodiment of the present disclosure, the cross-reconstruction network includes a first autoencoder and a second autoencoder, and the cross-reconstruction unit is specifically used to: input the attribute characteristics of the first user into the first autoencoder, encode the attribute characteristics using the autoencoder to obtain target encoded characteristics; and decode the target encoded characteristics using the second autodecoder to obtain the behavioral characteristics of the first user.

[0113] In one embodiment of the present disclosure, the prediction unit is specifically used to: input the attribute characteristics, behavioral characteristics and characteristics of the first user and the media information into an index prediction network to obtain a predicted value of a target index of the media information, wherein the index prediction network includes a fully connected layer, and the target index includes at least one of the following: click-through rate, conversion rate, and click-to-conversion rate.

[0114] In one embodiment of the present disclosure, the recommendation unit is specifically configured to: sort each piece of media information in the to-be-recommended media information set according to a predicted value of a target index of each piece of media information; and recommend the media information according to the sorting result.

[0115] In one embodiment of the present disclosure, the device also includes a first training unit, which is used to: obtain attribute characteristics and behavior characteristics of a second user, wherein the second user is a user with behavioral characteristics; and train the cross-reconstruction network using the attribute characteristics and behavior characteristics of the second user.

[0116] In one embodiment of the present disclosure, the cross-reconstruction network includes a first encoder, a second encoder, a first decoder and a second decoder, wherein the first encoder is used to encode an input attribute feature, the second encoder is used to encode an input behavior feature, the first decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, and the second decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, and the training loss of the cross-reconstruction network includes: a first loss, which is used to indicate the difference between the result obtained by the first decoder decoding the first encoding feature and the attribute feature of the second user, and the difference between the result obtained by the second decoder decoding the first encoding feature and the behavior feature of the second user; wherein the first encoding feature is obtained by the first encoder encoding the attribute feature of the second user; a second loss, which is used to indicate the difference between the result obtained by the first decoder decoding the second encoding feature and the attribute feature of the second user, and the difference between the result obtained by the second decoder decoding the second encoding feature and the behavior feature of the second user; wherein the second encoding feature is obtained by the second encoder encoding the behavior feature of the user.

[0117] In one embodiment of the present disclosure, the training loss of the cross reconstruction network also includes: a third loss, used to indicate the average difference between a first mapping result and a second mapping result, wherein the first mapping result is obtained by mapping the first encoding feature to a reproducing kernel Hilbert space, and the second mapping result is obtained by mapping the second encoding feature to a reproducing kernel Hilbert space.

[0118] In one embodiment of the present disclosure, the cross-reconstruction network also includes: a first attention mechanism network, used to map the embedding vector contained in the attribute features of the target user into a hidden attribute vector, and replace the attribute features with the hidden attribute vector, wherein the target user includes a first user and a second user; a second attention mechanism network, used to map the embedding vector contained in the behavior features of the second user into a hidden behavior vector, and replace the behavior features with the hidden behavior vector.

[0119] In one embodiment of the present disclosure, the apparatus further includes a joint training unit for jointly training the cross reconstruction network and the exponential prediction network.

[0120] It should be noted that, although several units / modules or sub-units / modules of the recommendation device are mentioned in the above detailed description, such division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above may be embodied in one unit / module. Conversely, the features and functions of one unit / module described above may be further divided to be embodied by multiple units / modules.

[0121] The present disclosure also provides a computer-readable storage medium. Figure 4 As shown, the storage medium stores a computer program 401, and when the computer program 401 is executed by a processor, the recommended method of any embodiment of the present disclosure can be executed.

[0122] An embodiment of the present disclosure further provides a computing device, which may include a memory and a processor, wherein the memory is used to store computer instructions executable on the processor, and the processor is used to implement the recommended method of any embodiment of the present disclosure when executing the computer instructions.

[0123] Figure 5 An example of a structure of the computing device is shown in FIG. Figure 5 As shown, the computing device 50 may include, but is not limited to: a processor 51, a memory 52, and a bus 53 connecting different system components (including the memory 52 and the processor 51).

[0124] The memory 52 stores computer instructions, which can be executed by the processor 51, so that the processor 51 can execute the recommended method of any embodiment of the present disclosure. The memory 52 may include a random access memory unit RAM 521, a cache memory unit 522 and / or a read-only memory unit ROM 523. The memory 52 may also include: a program tool 525 having a set of program modules 524, the program modules 524 include but are not limited to: an operating system, one or more application programs, other program modules and program data, and one or more combinations of these program modules may include the implementation of a network environment.

[0125] The bus 53 may include, for example, a data bus, an address bus, and a control bus. The computing device 50 may also communicate with an external device 55 through an I / O interface 54. The external device 55 may be, for example, a keyboard, a Bluetooth device, etc. The computing device 50 may also communicate with one or more networks through a network adapter 56. For example, the network may be a local area network, a wide area network, a public network, etc. Figure 5 As shown, the network adapter 56 can also communicate with other modules of the computing device 50 via the bus 53 .

[0126] In addition, although the operations of the disclosed method are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0127] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A recommendation method, characterized in that: include: Acquire attribute characteristics of a first user, wherein the first user is a user without behavioral characteristics; Inputting the attribute characteristics of the first user into a cross reconstruction network, cross-processing the attribute characteristics of the first user using the cross reconstruction network to obtain the behavior characteristics of the first user; the cross reconstruction network is used to predict the behavior characteristics of the first user based on the modal transfer relationship between user behavior and user attributes and the attribute characteristics of the first user; the cross processing refers to encoding the attribute characteristics using one of the dual autoencoders and decoding the encoded characteristics of the attribute characteristics using the other autoencoder; Obtaining a score for the media information according to the attribute characteristics and behavior characteristics of the first user and the characteristics of the media information; The first user is recommended according to the score of the media information.

2. The method according to claim 1, characterized in that The cross-reconstruction network includes a first autoencoder and a second autoencoder, and cross-processing the attribute features of the first user using the cross-reconstruction network to obtain the behavior features of the first user includes: Inputting the attribute features of the first user into the first autoencoder, and encoding the attribute features using the first autoencoder to obtain target encoded features; The target encoded features are decoded using the second autoencoder to obtain behavioral features of the first user.

3. The method according to claim 1, characterized in that The obtaining a score of the media information according to the attribute characteristics and behavior characteristics of the first user and the characteristics of the media information includes: The attribute characteristics, behavior characteristics and characteristics of the first user and the characteristics of the media information are input into an index prediction network to obtain a predicted value of a target index of the media information, wherein the index prediction network includes a fully connected layer, and the target index includes at least one of the following: click-through rate, conversion rate, and click-to-conversion rate.

4. The method according to claim 3, characterized in that The recommending the first user according to the score of the media information includes: sorting each piece of media information in the set of media information to be recommended according to the predicted value of the target index of each piece of media information; The media information is recommended according to the sorting result.

5. The method according to any one of claims 1 to 4, characterized in that: The method further comprises: Acquire attribute characteristics and behavior characteristics of a second user, wherein the second user is a user with behavior characteristics; The cross reconstruction network is trained using the attribute characteristics and behavior characteristics of the second user.

6. The method according to claim 5, characterized in that The cross reconstruction network includes a first encoder, a second encoder, a first decoder and a second decoder, wherein the first encoder is used to encode the input attribute features, the second encoder is used to encode the input behavior features, the first decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, the second decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, and the training loss of the cross reconstruction network includes: A first loss is used to indicate a difference between a result obtained by the first decoder decoding the first coding feature and an attribute feature of the second user, and a difference between a result obtained by the second decoder decoding the first coding feature and a behavior feature of the second user; wherein the first coding feature is obtained by the first encoder encoding the attribute feature of the second user; A second loss is used to indicate a difference between a result obtained by the first decoder decoding the second coding feature and the attribute characteristics of the second user, and a difference between a result obtained by the second decoder decoding the second coding feature and the behavior characteristics of the second user; wherein the second coding feature is obtained by the second encoder encoding the behavior characteristics of the second user.

7. The method according to claim 6, characterized in that The training loss of the cross reconstruction network also includes: a third loss, used to indicate the average difference between a first mapping result and a second mapping result, wherein the first mapping result is obtained by mapping the first encoding feature to a reproducing kernel Hilbert space, and the second mapping result is obtained by mapping the second encoding feature to a reproducing kernel Hilbert space.

8. The method according to claim 6 or 7, characterized in that: The cross-reconstruction network also includes: A first attention mechanism network, used to map an embedding vector included in an attribute feature of a target user into a hidden attribute vector, and replace the attribute feature with the hidden attribute vector, wherein the target user includes a first user and a second user; The second attention mechanism network is used to map the embedding vector contained in the behavior feature of the second user into a hidden behavior vector, and replace the behavior feature with the hidden behavior vector.

9. The method according to claim 3, characterized in that: The method also includes: jointly training the cross reconstruction network and the exponential prediction network.

10. A recommendation device, characterized in that: include: An acquiring unit, configured to acquire attribute characteristics of a first user, wherein the first user is a user without behavioral characteristics; a cross reconstruction unit, configured to input the attribute characteristics of the first user into a cross reconstruction network, and use the cross reconstruction network to perform cross processing on the attribute characteristics of the first user to obtain the behavior characteristics of the first user; the cross reconstruction network is configured to predict the behavior characteristics of the first user according to the modal transfer relationship between the user behavior and the user attribute and the attribute characteristics of the first user; the cross processing refers to encoding the attribute characteristics by using one of the dual autoencoders, and decoding the encoded characteristics of the attribute characteristics by using the other autoencoder; A prediction unit, configured to obtain a score of the media information according to the attribute characteristics and behavior characteristics of the first user and the characteristics of the media information; A recommendation unit is used to recommend the first user according to the score of the media information.

11. The device according to claim 10, characterized in that The cross reconstruction network includes a first autoencoder and a second autoencoder, and the cross reconstruction unit is specifically used for: Inputting the attribute features of the first user into the first autoencoder, and encoding the attribute features using the autoencoder to obtain target encoded features; The target encoded features are decoded using the second autoencoder to obtain behavioral features of the first user.

12. The device according to claim 10, characterized in that The prediction unit is specifically used for: The attribute characteristics, behavior characteristics and characteristics of the first user and the characteristics of the media information are input into an index prediction network to obtain a predicted value of a target index of the media information, wherein the index prediction network includes a fully connected layer, and the target index includes at least one of the following: click-through rate, conversion rate, and click-to-conversion rate.

13. The device according to claim 12, characterized in that The recommendation unit is specifically used for: sorting each piece of media information in the set of media information to be recommended according to the predicted value of the target index of each piece of media information; The media information is recommended according to the sorting result.

14. The device according to any one of claims 10 to 13, characterized in that Also included is a first training unit, for: Acquire attribute characteristics and behavior characteristics of a second user, wherein the second user is a user with behavior characteristics; The cross reconstruction network is trained using the attribute characteristics and behavior characteristics of the second user.

15. The device according to claim 14, characterized in that The cross reconstruction network includes a first encoder, a second encoder, a first decoder and a second decoder, wherein the first encoder is used to encode the input attribute features, the second encoder is used to encode the input behavior features, the first decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, the second decoder is used to decode the encoding result output by the first encoder and the encoding result output by the second encoder, and the training loss of the cross reconstruction network includes: A first loss is used to indicate a difference between a result obtained by the first decoder decoding the first coding feature and an attribute feature of the second user, and a difference between a result obtained by the second decoder decoding the first coding feature and a behavior feature of the second user; wherein the first coding feature is obtained by the first encoder encoding the attribute feature of the second user; The second loss is used to indicate the difference between the result obtained by the first decoder decoding the second coding feature and the attribute characteristics of the second user, and the difference between the result obtained by the second decoder decoding the second coding feature and the behavior characteristics of the second user; wherein the second coding feature is obtained by the second encoder encoding the behavior characteristics of the user.

16. The device according to claim 15, characterized in that The training loss of the cross reconstruction network also includes: a third loss, used to indicate the average difference between a first mapping result and a second mapping result, wherein the first mapping result is obtained by mapping the first encoding feature to a reproducing kernel Hilbert space, and the second mapping result is obtained by mapping the second encoding feature to a reproducing kernel Hilbert space.

17. The device according to claim 15 or 16, characterized in that The cross-reconstruction network also includes: A first attention mechanism network, used to map an embedding vector included in an attribute feature of a target user into a hidden attribute vector, and replace the attribute feature with the hidden attribute vector, wherein the target user includes a first user and a second user; The second attention mechanism network is used to map the embedding vector contained in the behavior feature of the second user into a hidden behavior vector, and replace the behavior feature with the hidden behavior vector.

18. The device according to claim 12, characterized in that It also includes a joint training unit for jointly training the cross reconstruction network and the exponential prediction network.

19. A computing device, characterized in that include: processor; as well as A memory, wherein computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by the processor, the recommendation method according to any one of claims 1 to 9 is implemented.

20. A computer-readable storage medium, characterized in that: The invention comprises a computer program, which implements the recommendation method according to any one of claims 1 to 9 when being executed by a processor.

Images