CN115114521A

CN115114521A - Training method, using method, device, equipment and medium of feature reconstruction model

Info

Publication number: CN115114521A
Application number: CN202210624665.4A
Authority: CN
Inventors: 白冰
Original assignee: Tencent Technology Beijing Co Ltd
Current assignee: Tencent Technology Beijing Co Ltd
Priority date: 2022-06-02
Filing date: 2022-06-02
Publication date: 2022-09-27

Abstract

The application discloses a training method, a using method, a device, equipment and a medium of a feature reconstruction model, and belongs to the technical field of artificial intelligence. The method comprises the following steps: acquiring a first identification characteristic of a first account, and acquiring a second identification characteristic and an auxiliary characteristic of a second account; coding the first identification feature to obtain a first reconstruction identification feature, coding the second identification feature to obtain a second reconstruction identification feature, and coding the auxiliary feature to obtain a reconstruction auxiliary feature; predicting account types to which the first reconstruction passing identification feature and the second reconstruction passing identification feature belong to obtain a prediction classification result; and training the characteristic reconstruction model according to the contrast error and the confrontation error. The method and the device reduce the difference between accounts with the account types of the first type and the second type by training the feature reconstruction model; the problem of difficulty in clustering caused by information difference of accounts among different types is solved.

Description

Training method, using method, device, equipment and medium of feature reconstruction model

Technical Field

The application relates to the technical field of artificial intelligence, in particular to a training method, a using method, a device, equipment and a medium of a feature reconstruction model.

Background

With the development of internet technology, the demand for clustering different users and performing personalized services according to different clustering results is increasing day by day.

In the related art, the clustering result of the user is usually obtained by extracting the behavior information of the user in the application program, such as browsing information of the user on multimedia information and usage information of the user on the application program, and clustering the user according to the behavior information.

However, the method depends heavily on the behavior information of the user in the application program, and for the user who uses the application program for the first time or the user who has little behavior information in the application program, the classification is often difficult, and how to improve the clustering result is an urgent problem to be solved.

Disclosure of Invention

The application provides a training method, a using method, a device, equipment and a medium of a feature reconstruction model, and the technical scheme is as follows:

according to an aspect of the present application, there is provided a training method of a feature reconstruction model, the method including:

acquiring a first identification feature of a first account, and acquiring a second identification feature and an auxiliary feature of a second account, wherein the first identification feature comprises a feature representation of first identification information of the first account, the second identification feature comprises a feature representation of second identification information of the second account, and the auxiliary feature comprises a feature representation of auxiliary information of the second account; the account type of the first account is a first type, the account type of the second account is a second type, and the account in the first type has no corresponding auxiliary information;

coding the first identification feature to obtain a first reconstruction identification feature, coding the second identification feature to obtain a second reconstruction identification feature, and coding the auxiliary feature to obtain a reconstruction auxiliary feature;

predicting account types to which the first reconstruction passing identification feature and the second reconstruction passing identification feature belong to obtain a prediction classification result;

training the feature reconstruction model according to a comparison error and a confrontation error to obtain a trained feature reconstruction model, wherein the comparison error comprises an error between the second reconstruction identification feature and the reconstruction auxiliary feature, the confrontation error comprises an error between the prediction classification result and an actual classification result, and the actual classification result is an account type corresponding to the first identification feature and the second identification feature.

According to another aspect of the present application, there is provided a method for using a feature reconstruction model, the feature reconstruction model being trained according to the above training method for the feature reconstruction model, the method including:

acquiring a passing identification characteristic of at least one third account, wherein the passing identification characteristic comprises a characteristic representation of passing identification information of the third account;

coding the pass identification features based on the feature reconstruction model to obtain reconstructed pass identification features;

and clustering the reconstructed passing identification characteristics to obtain a clustering result of the third account.

According to another aspect of the present application, there is provided an information recommendation method, the method including:

acquiring application characteristics of a plurality of target application accounts, wherein the target application accounts comprise at least one fourth account with an inactive account type and at least one fifth account with an active account type, the application characteristics comprise characteristic representation of application information of the target application accounts, and the application information comprises installation and uninstallation information of the target application accounts on at least one application;

coding the application characteristics based on a characteristic reconstruction model to obtain reconstructed application characteristics, wherein the characteristic reconstruction model is obtained by training according to a training method of the characteristic reconstruction model;

clustering the reconstruction application characteristics to obtain a clustering result of the target application account;

and determining information recommendation information of the fourth account belonging to the same clustering result as the fifth account based on the information of the fifth account, wherein the information comprises information browsing information of the fifth account in the target application.

According to another aspect of the present application, there is provided a training apparatus for a feature reconstruction model, the apparatus including:

the acquisition module is used for acquiring a first identification feature of a first account, and acquiring a second identification feature and an auxiliary feature of a second account, wherein the first identification feature comprises a feature representation of first identification information of the first account, the second identification feature comprises a feature representation of second identification information of the second account, and the auxiliary feature comprises a feature representation of auxiliary information of the second account; the account type of the first account is a first type, the account type of the second account is a second type, and the account in the first type has no corresponding auxiliary information;

the encoding module is used for encoding the first identification characteristics to obtain first reconstruction identification characteristics, encoding the second identification characteristics to obtain second reconstruction identification characteristics, and encoding the auxiliary characteristics to obtain reconstruction auxiliary characteristics;

the prediction module is used for predicting account types to which the first reconstruction passing identification feature and the second reconstruction passing identification feature belong to obtain a prediction classification result;

the training module is used for training the feature reconstruction model according to a comparison error and a confrontation error to obtain the trained feature reconstruction model, the comparison error comprises an error between the second reconstruction recognition feature and the reconstruction auxiliary feature, the confrontation error comprises an error between the prediction classification result and an actual classification result, and the actual classification result is an account type corresponding to the first recognition feature and the second recognition feature.

In an alternative design of the present application, the feature reconstruction model includes a contrast learning encoder;

the encoding module is further configured to:

calling the contrast learning encoder to encode the first pass identification feature to obtain a first reconstruction pass identification feature, calling the contrast learning encoder to encode the second pass identification feature to obtain a second reconstruction pass identification feature, and calling the contrast learning encoder to encode the auxiliary feature to obtain the reconstruction auxiliary feature.

In an alternative design of the present application, the feature reconstruction model further includes a general encoder and an auxiliary encoder;

the acquisition module is further configured to:

acquiring the first identification information of a first account, and acquiring the second identification information and the auxiliary information of a second account;

calling the identification encoder to encode the first identification information to obtain the first identification feature, calling the identification encoder to encode the second identification information to obtain the second identification feature, and calling the auxiliary encoder to encode the auxiliary information to obtain the auxiliary feature.

In an alternative design of the present application, the first identification information and the second identification information are multi-hot coded vectors, and the auxiliary information is continuous value vectors.

In an alternative design of the present application, the feature reconstruction model further includes a general decoder and an auxiliary decoder;

the device further comprises:

the decoding module is used for calling the pass-identification decoder to decode the first pass-identification feature to obtain first reconstruction pass-identification information, calling the pass-identification decoder to decode the second pass-identification feature to obtain second reconstruction pass-identification information, and calling the auxiliary decoder to decode the auxiliary feature to obtain reconstruction auxiliary information;

the training module is further configured to:

training the feature reconstruction model according to the comparison error, the confrontation error and the reconstruction error to obtain the trained feature reconstruction model;

wherein the reconstruction error comprises: at least one of an error between the first reconstruction identification information and the first identification information, an error between the second reconstruction identification information and the second identification information, and an error between the reconstruction assistance information and the assistance information.

In an alternative design of the present application, at least one of the contrast error, the countermeasure error, and the reconstruction error includes a cross-entropy loss.

In an alternative design of the present application, the feature reconstruction model includes a gradient inversion layer and an account classifier;

the prediction module is further to:

calling the gradient inversion layer to perform inversion processing on the first reconstruction passing identification feature and the second reconstruction passing identification feature to obtain a processed first reconstruction passing identification feature and a processed second reconstruction passing identification feature;

and calling the account classifier to predict the account types of the processed first reconstruction passing identification feature and the processed second reconstruction passing identification feature to obtain a prediction classification result.

In an optional design of the present application, the first identification information is first application information, the first application information includes installation and uninstallation information of the first account for the at least one application, the second identification information is second application information, the second application information includes installation and uninstallation information of the second account for the at least one application, the auxiliary information is information, and the information includes information browsing information of the second account in a target application;

the acquisition module is further configured to:

acquiring a first application characteristic of the first account, and acquiring a second application characteristic and an information characteristic of the second account, wherein the first application characteristic comprises a characteristic representation of the first application information of the first account, the second application characteristic comprises a characteristic representation of the second application information of the second account, and the information characteristic comprises a characteristic representation of the information of the second account;

the encoding module is further configured to:

coding the first application characteristic to obtain a first reconstruction application characteristic, coding the second application characteristic to obtain a second reconstruction application characteristic, and coding the information characteristic to obtain a reconstruction information characteristic;

the prediction module is further to:

and predicting the account types of the first reconstruction application characteristic and the second reconstruction application characteristic to obtain the prediction classification result.

According to another aspect of the present application, there is provided an apparatus for using a feature reconstruction model, the apparatus including:

the acquisition module is used for acquiring the identification passing characteristics of at least one third account, wherein the identification passing characteristics comprise characteristic representation of the identification passing information of the third account;

the coding module is used for coding the identification characteristics based on the characteristic reconstruction model to obtain reconstructed identification characteristics;

and the clustering module is used for clustering the reconstruction passing identification characteristics to obtain a clustering result of the third account.

In an alternative design of the present application, the apparatus further includes:

and the determining module is used for acquiring auxiliary information of the third account in a target time period and determining account characteristics of a target cluster according to the auxiliary information, wherein the target cluster comprises a cluster indicated by a cluster result of the third account.

In an alternative design of the present application, the determining module is further configured to:

and calling a sorting model to sort the account features to obtain recommendation information of the account to be recommended, wherein the account to be recommended belongs to the target cluster.

the determining module is configured to acquire high-frequency auxiliary information that the third account exceeds a target threshold within a target time period, and determine recommendation information of an account to be recommended according to the high-frequency auxiliary information, where the third account and the account to be recommended belong to a target cluster, and the target cluster includes a cluster indicated by a cluster result of the third account.

According to another aspect of the present application, there is provided an information recommendation apparatus, the apparatus including:

the system comprises an acquisition module, a processing module and a display module, wherein the acquisition module is used for acquiring application characteristics of a plurality of target application accounts, the target application accounts comprise at least one fourth account with an inactive account type and at least one fifth account with an active account type, the application characteristics comprise characteristic representation of application information of the target application accounts, and the application information comprises installation and uninstallation information of at least one application of the target application accounts;

the coding module is used for coding the application characteristics based on the characteristic reconstruction model to obtain reconstructed application characteristics;

the clustering module is used for clustering the reconstruction application characteristics to obtain a clustering result of the target application account;

and the determining module is used for determining information recommendation information of the fourth account belonging to the same clustering result as the fifth account based on the information of the fifth account, wherein the information comprises information browsing information of the fifth account in target application.

According to another aspect of the present application, there is provided a computer apparatus comprising a processor and a memory having stored therein at least one instruction, at least one program, set of codes, or set of instructions that is loaded and executed by the processor to implement a method of reconstructing a model of features and/or a method of use and/or a method of information recommendation 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, which is loaded and executed by a processor to implement a method of reconstructing a model of features and/or a method of using and/or a method of information recommendation as described above.

According to another aspect of the present application, there is provided a computer program product comprising computer instructions stored in a computer-readable storage medium, which are read and executed by a processor to implement the method of feature reconstruction model and/or the method of use and/or the method of information recommendation described above.

The beneficial effect that technical scheme that this application provided brought includes at least:

the characteristics are reconstructed through coding processing, so that the difference between the identification information and the auxiliary information is reduced, a prediction classification result is obtained through prediction, and the difference between accounts with the account types of the first type and the second type is reduced; the characteristic reconstruction model is trained according to the comparison error and the countererror, so that the accounts with the aligned account types of the first type and the second type are realized, the influence of auxiliary information on second identification information of the second account in the clustering process is eliminated, the effect of clustering the accounts is ensured, and the problem of difficulty in clustering caused by the information difference of the accounts among different types is solved.

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 computer system provided in an exemplary embodiment of the present application;

FIG. 2 is a schematic diagram of a training method for a feature reconstruction model provided in an exemplary embodiment of the present application;

FIG. 3 is a schematic diagram of a method of using a feature reconstruction model provided by an exemplary embodiment of the present application;

FIG. 4 is a flow chart of a method for training a feature reconstruction model provided in an exemplary embodiment of the present application;

FIG. 5 is a flow chart of a method for training a feature reconstruction model provided in an exemplary embodiment of the present application;

FIG. 6 is a flow chart of a method for training a feature reconstruction model provided in an exemplary embodiment of the present application;

FIG. 7 is a flowchart of a method for training a feature reconstruction model provided by an exemplary embodiment of the present application;

FIG. 8 is a flow chart of a method for training a feature reconstruction model provided in an exemplary embodiment of the present application;

FIG. 9 is a flowchart of a training method for a feature reconstruction model provided in an exemplary embodiment of the present application;

FIG. 10 is a flow chart of a method of using a feature reconstruction model provided by an exemplary embodiment of the present application;

FIG. 11 is a flowchart of an information recommendation method provided in an exemplary embodiment of the present application;

FIG. 12 is a flowchart of an information recommendation method provided in an exemplary embodiment of the present application;

FIG. 13 is an interface diagram of recommendation information provided by an exemplary embodiment of the present application;

FIG. 14 is a flowchart of an information recommendation method provided in an exemplary embodiment of the present application;

FIG. 15 is a diagram of an account clustering model provided by an exemplary embodiment of the present application;

FIG. 16 is a block diagram of a training apparatus for feature reconstruction models according to an exemplary embodiment of the present application;

FIG. 17 is a block diagram of an apparatus for using a feature reconstruction model according to an exemplary embodiment of the present application;

FIG. 18 is a block diagram illustrating an information recommendation device according to an exemplary embodiment of the present application;

fig. 19 is a block diagram of a server 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.

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

The terminology used in the present disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. As used in this disclosure and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

It should be noted that, the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) referred to in the present application are information and data authorized by the user or sufficiently authorized by each party, and the collection, use and processing of the related data need to comply with the relevant laws and regulations and standards of the relevant country and region. For example, the first identification information of the first account, the second identification information of the second account, and the auxiliary information referred to in this application are all obtained under the condition of sufficient authorization of the user.

It will be understood that, although the terms first, second, etc. may be used herein to describe various information, these information should not be limited by these terms. These terms are only used to distinguish one type of information from another. For example, a first parameter may also be referred to as a second parameter, and similarly, a second parameter may also be referred to as a first parameter, without departing from the scope of the present disclosure. The word "if" as used herein may be interpreted as "at … …" or "when … …" or "in response to a determination", depending on the context.

FIG. 1 illustrates a schematic diagram of a computer system provided by one embodiment of the present application. The computer system may implement a training method and/or a system architecture using the method or system architecture into a feature reconstruction model. The computer system may include: a terminal 100 and a server 200. The terminal 100 may be an electronic device such as a mobile phone, a tablet Computer, a vehicle-mounted terminal (car machine), a wearable device, a PC (Personal Computer), an unmanned terminal, and the like. The terminal 100 may have a client installed therein for running a target application, which may be a training and/or using application of the feature reconstruction model, or may be another application provided with a training and/or using function of the feature reconstruction model, and this application is not limited thereto. The form of the target Application is not limited in the present Application, and may include, but is not limited to, an App (Application program) installed in the terminal 100, an applet, and the like, and may be a web page form. The server 200 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing cloud computing services. The server 200 may be a background server of the target application program, and is configured to provide a background service for a client of the target application program.

According to the training method and/or the using method of the feature reconstruction model provided by the embodiment of the application, the execution subject of each step can be computer equipment, and the computer equipment refers to electronic equipment with data calculation, processing and storage capabilities. Taking the embodiment environment shown in fig. 1 as an example, the terminal 100 may perform the training method and/or the using method of the feature reconstruction model (for example, a client installed in the terminal 100 and running a target application program performs the training method and/or the using method of the feature reconstruction model), the server 200 may perform the training method and/or the using method of the feature reconstruction model, or the terminal 100 and the server 200 cooperate with each other to perform the training method and/or the using method of the feature reconstruction model, which is not limited in this application.

In addition, the technical scheme of the application can be combined with the block chain technology. For example, some of the data involved in the training and/or use methods of the feature reconstruction model disclosed herein may be stored on a blockchain. The terminal 100 and the server 200 may communicate with each other through a network, such as a wired or wireless network.

In the related art, it is necessary to perform clustering processing on different accounts and perform personalized service according to different clustering results. Because the account types are divided into an active type and an inactive type, the account of different account types has completely different behavior information in the application program, and the account of the active type and the account of the inactive type have field difference.

Such as: taking the information indicating the browsing behavior of the account on the multimedia information in the application program as an example, the account with the account type being active type corresponds to the information, and the account with the account type being inactive type does not have the corresponding information. Further, the application information of the account is used for indicating installation and uninstallation information of at least one application; the information corresponding to the active type account can affect the application information; for example, accounts with high frequency of multimedia browsing behavior have more people applications installed.

In the process of clustering accounts according to the behavior information of the accounts in the application program, because the field difference exists between the active type accounts and the inactive type accounts, the clustering result of the accounts is poor; in one implementation, a clustering process in the related art clusters accounts of the active type and the inactive type into different cluster clusters, respectively.

In the application, aiming at the field difference between the accounts of the active type and the inactive type, the difference between the information and the application information of the account of the active type is reduced; and by reducing the difference between the application information of the account of the active type and the application information of the account of the inactive type, the influence of the field difference between the accounts of the active type and the inactive type on account clustering is eliminated, and the effect of clustering the accounts is ensured.

Next, a feature reconstruction model in the present application is introduced:

fig. 2 is a schematic diagram illustrating a training method of a feature reconstruction model according to an embodiment of the present application.

The feature reconstruction model 330 includes: an encoding network 330a and a classification network 330 b.

The encoding network 330a includes: a general encoder 332a, an auxiliary encoder 332b, and a contrast learning encoder 332 c; illustratively, in this embodiment, the encoding network 330a further includes: a general decoder 338a and an auxiliary decoder 338 b.

The classification network 330b includes: gradient inversion layer 334 and account number sorter 336.

Illustratively, the first application information 312 of the first account 310 is acquired, and the second application information 322 and the information 324 of the second account 320 are acquired.

Illustratively, the first application information 312 includes installation and uninstallation information of the first account 310 for the at least one application, the second application information 322 includes installation and uninstallation information of the second account 320 for the at least one application, and the information 324 includes information browsing information of the second account 320 in the target application.

The account type of the first account 310 is an inactive type, and the account type of the second account 320 is an active type; the account of the inactive type has no corresponding information.

Calling a recognition encoder 332a to perform encoding processing on the first application information 312 to obtain a first application characteristic 312 a; the second application information 322 is encoded by the generic identification encoder 332a, and the second application characteristics 322a are obtained.

The auxiliary encoder 332b is invoked to encode the information message 324, so as to obtain the information feature 324 a.

Invoking a contrast learning encoder 332c to perform encoding processing on the first application feature 312a to obtain a first reconstructed application feature 312 b; invoking a contrast learning encoder 332c to perform encoding processing on the second application feature 322a to obtain a second reconstructed application feature 322 b; the contrast learning encoder 332c is invoked to encode the information feature 324a to obtain the reconstructed information feature 324 b.

Exemplarily, the present embodiment further includes:

invoking a recognition decoder 338a to perform encoding processing on the first application feature 312a to obtain first reconstructed application information 312 d; the recognition decoder 338a is invoked to encode the second application characteristic 322a, resulting in the second reconstructed application information 322 d.

The auxiliary decoder 338b is invoked to encode the information feature 324a to obtain the reconstructed information 324 d.

The classification network 330b is invoked to predict the account types to which the first reconstruction application feature 312b and the second reconstruction application feature 322b belong.

Specifically, the gradient inversion layer 334 is invoked to perform inversion processing on the first reconstruction application feature 312b and the second reconstruction application feature 322b, so as to obtain a processed first reconstruction application feature and a processed second reconstruction application feature; the account classifier 336 is invoked to predict the account types of the processed first reconstructed application feature and the processed second reconstructed application feature, so as to obtain a first predicted classification result 312c of the first reconstructed application feature 312b and a second predicted classification result 322c of the second reconstructed application feature 322 b.

Based on the comparison error between the second reconstructed application feature 322b and the reconstructed information feature 324b, the countermeasure error between the first predicted classification result 312c, the second predicted classification result 322c and the actual classification result, and the reconstruction error between the first application information 312 and the first reconstructed application information 312d, the second application information 322 and the second reconstructed application information 322d, the information 324 and the reconstructed information 324 d; and training the feature reconstruction model 330 to obtain a trained feature reconstruction model.

The actual classification result is the account types actually corresponding to the first application information 312 and the second application information 322.

Fig. 3 is a schematic diagram illustrating a method for using a feature reconstruction model according to an embodiment of the present application.

The trained feature reconstruction model 420 includes an encoding network 420 a; specifically, the encoding network 420a includes a trained consensus encoder 422a and a trained contrast learning encoder 422 b.

Acquiring application information 412 of the at least one third account, wherein the application information 412 includes installation and uninstallation information of the at least one application by the third account.

And calling the trained feature reconstruction model 420 to encode the application information 412 to obtain a reconstructed application feature 412 b.

Specifically, the trained recognition encoder 422a is called to encode the application information 412, so as to obtain an application feature 412 a; and calling the trained contrast learning coder 422b to code the application features 412a to obtain reconstructed application features 412 b.

And clustering the reconstructed application features 412b by calling the clustering network 432 to obtain a clustering result 412c of at least one third account.

Illustratively, the clustering result 412c of the at least one third account is used to indicate at least one cluster, and at least one account with the same account characteristics is included in one cluster. Such as: the first cluster corresponds to at least one account with cartoon preference, and the second cluster corresponds to at least one account with news preference; it should be noted that the description of the cluster is only an exemplary introduction, the accounts corresponding to the cluster have the same account characteristics, and the account characteristics may directly indicate that there is a preference characteristic, or may be only an abstract account characteristic, and do not indicate the preference content of the account.

In order to improve the feature reconstruction effect of the feature reconstruction model, the feature reconstruction model needs to be trained, and then, the training method of the feature reconstruction model will be described through the following embodiments.

Fig. 4 shows a flowchart of a training method of a feature reconstruction model according to an exemplary embodiment of the present application. The method may be performed by a computer device. The method comprises the following steps:

step 510: acquiring a first identification characteristic of a first account, and acquiring a second identification characteristic and an auxiliary characteristic of a second account;

the first identification characteristic comprises a characteristic representation of the first identification information of the first account. Illustratively, the first identification information is attribute information of the first account, or information that has a binding relationship with the first account; such as information determined or obtained when registering the first account. In one example, each first account corresponds to corresponding first identification information, and the first account also corresponds to the first identification information when the first account is not used or accessed.

The second identification characteristic comprises a characteristic representation of the second identification information of the second account and the secondary characteristic comprises a characteristic representation of the secondary information of the second account. Illustratively, the second identification information is attribute information of the second account, or information that has a binding relationship with the second account. Illustratively, the auxiliary information is information generated by the second account in the application process; such as information determined or obtained when using or accessing the second account.

Exemplary representations of the feature representation include, but are not limited to: at least one of an eigenvector, an eigenvector matrix, an eigenvalue, or bit information.

The account type of the first account is a first type, and the account type of the second account is a second type; the information abundance degree of the account in the first type is lower than that of the account in the second type, and the account in the first type does not have corresponding auxiliary information. Illustratively, the first type is an inactive type and the second type is an active type.

For example, accounts in the first type typically do not have corresponding auxiliary information due to fewer accounts than the second account; but does not exclude the case that the account in the first type does not have the corresponding auxiliary information because the corresponding auxiliary information cannot be acquired.

Step 520: coding the first identification feature to obtain a first reconstruction identification feature, coding the second identification feature to obtain a second reconstruction identification feature, and coding the auxiliary feature to obtain a reconstruction auxiliary feature;

for example, the encoding process may be realized by performing prediction through a feature reconstruction model, or may be realized by performing statistical calculation through the feature reconstruction model, which is not limited in this embodiment.

For example, the first reconstructed generic identification feature and the first generic identification feature may be represented in the same manner or different manners, including but not limited to: at least one of an eigenvector, an eigenvector matrix, an eigenvalue, or bit information. Similarly, the second identifying feature, the second reconstruction identifying feature, the assistant feature and the reconstruction assistant feature may be represented in the same manner or in different manners.

It should be noted that the encoding processes in this step may be executed simultaneously, or may be executed sequentially in any order; that is, the present embodiment does not provide any restrictive provisions for the timing relationship of the encoding process performed by the first identifying characteristic, the second identifying characteristic, and the assistant characteristic.

Step 530: predicting account types to which the first reconstruction passing identification feature and the second reconstruction passing identification feature belong to obtain a prediction classification result;

illustratively, the predicted classification result indicates account types to which the predicted first and second reconstruction recognition features belong. Illustratively, the predictive classification result includes probability information of the first type and/or the second type.

Step 540: training the feature reconstruction model according to the comparison error and the confrontation error to obtain a trained feature reconstruction model;

illustratively, the contrast error includes an error between the second reconstruction pass identification feature and the reconstruction assisting feature.

The countermeasure error comprises an error between the predicted classification result and an actual classification result, and the actual classification result is an account type actually corresponding to the first identification feature and the second identification feature.

Illustratively, the countermeasure error is used to guide the trained feature reconstruction model to not predict the actual classification result. Such as: the trained feature reconstruction model cannot accurately predict account types to which the first reconstruction passing identification feature and the second reconstruction passing identification feature belong.

Illustratively, training the feature reconstruction model by comparing the error with the countererror reduces the difference between the description of the second identification information and the auxiliary information of the second account by the feature representation, and reduces the difference between the description of the first identification information of the first account and the description of the second identification information of the second account by the feature representation, i.e. aligns the accounts with the first and second account types.

In summary, the method provided in this embodiment reduces the difference between the general identification information and the auxiliary information by encoding and processing the reconstruction features, obtains the prediction classification result by prediction, and reduces the difference between accounts with the account types of the first type and the second type; the characteristic reconstruction model is trained according to the comparison error and the countererror, so that the accounts with the aligned account types of the first type and the second type are realized, the influence of auxiliary information on second identification information of the second account in the clustering process is eliminated, the effect of clustering the accounts is ensured, and the problem of difficulty in clustering caused by the information difference of the accounts among different types is solved.

Fig. 5 is a flowchart illustrating a training method of a feature reconstruction model according to an exemplary embodiment of the present application. The method may be performed by a computer device. That is, in the embodiment shown in fig. 4, step 520 may be implemented as step 522:

step 522: calling a contrast learning encoder to encode the first pass identification feature to obtain a first reconstruction pass identification feature, calling the contrast learning encoder to encode the second pass identification feature to obtain a second reconstruction pass identification feature, and calling the contrast learning encoder to encode the auxiliary feature to obtain a reconstruction auxiliary feature;

illustratively, the feature reconstruction model includes a contrast learning encoder. Wherein the contrast learning encoder includes, but is not limited to, at least one of the following networks: convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long-Short Term Memory Networks (LSTM).

Illustratively, the encoding processes of the contrast learning encoder are independent of each other; specifically, the method comprises the following steps:

h ^(v) ＝f _con (z ^(v) ；θ _c )；

wherein f is _con Representing contrast learning encoders, theta _c Parameters representing a contrast learning encoder; z is a radical of ^(v) Represents a feature representation, h ^(v) The representation features represent corresponding reconstruction features.

Such as: in the case where v is 1, z ⁽¹⁾ Representing a first or a second communication characteristic, h ⁽¹⁾ Representing the corresponding first or second reconstructed generic features.

In the case where v is 2, z ⁽²⁾ Denotes an assist feature, h ⁽²⁾ Representing the corresponding reconstruction assisting feature.

In an alternative implementation, at least one of the first reconstructed common feature, the second reconstructed common feature and the reconstructed assist feature is a continuous-valued vector. In one implementation, a continuous-value Vector is also referred to as a Dense Vector (Dense Vector); it will be appreciated that the dimensions of the first, second and reconstruction assisting features are typically the same, but that the presence of different dimensions is not excluded.

In summary, the method provided in this embodiment expands the construction method of the feature reconstruction model by calling the contrast learning encoder to perform encoding processing, and reduces the difference between the general identification information and the auxiliary information by representing the features obtained by encoding with the contrast learning encoder; the method and the device ensure the effect of clustering accounts and avoid the problem of difficulty in clustering due to the information difference of accounts of different types.

Next, details of the comparison error are presented:

illustratively, the contrast error is used to describe an error between the second reconstruction pass identification feature and the reconstruction assisting feature; it is understood that the contrast error includes, but is not limited to, at least one of: cross Entropy Loss function (Cross-Encopy Loss), 0-1 Loss function (Zero-One Loss), and Dis Loss function (Dice Loss).

In one example, the contrast error includes:

wherein L is _con Indicating contrast error, U _a The number of the second account number is shown, the temperature over-parameter of contrast learning is shown by tau,

to represent

And

the cosine similarity between the two signals is determined,

a second reconfiguration generic feature representing an ith second account,

indicating the reconstruction assistant feature of the jth second account.

It is noted that in one example, the contrast error is used to describe the error between the second reconstruction recognition feature and the reconstruction assisting feature; training the feature reconstruction model through the comparison error, adjusting coding parameters when coding processing is carried out to obtain a second reconstruction passing identification feature and a reconstruction auxiliary feature, optionally, obtaining the second reconstruction passing identification feature through calling a comparison learning coder to code the second passing identification feature, and calling the comparison learning coder to code the auxiliary feature to obtain the reconstruction auxiliary feature.

Illustratively, the contrast error reduces the difference between the second reconstruction recognition feature and the reconstruction assisting feature, and the feature reconstruction model is trained; the comparison error is used for eliminating the field difference between the auxiliary information of the second account and the identification information of the second account, and the comparison error is used for eliminating the influence of the auxiliary information of the second account on the second identification information of the second account; the comparison error aligns the second identification information and the auxiliary information.

Furthermore, the feature reconstruction model is trained through comparison errors, and encoding parameters of a comparison learning encoder are adjusted; the comparison learning encoder carries out encoding processing on the first pass identification feature to obtain a first reconstruction pass identification feature; and the contrast learning encoder encodes the second identification feature to obtain a second reconstructed identification feature. The field difference between the auxiliary information of the second account and the general identification information of the second account is eliminated by comparing errors, and a foundation is laid for reducing the field information carried in the first reconstruction general identification characteristic and the second reconstruction general identification characteristic and eliminating the field difference caused by the difference of the account types between the first account and the second account by predicting the account types.

Further, for the first account, the account type of the first account is an inactive type, and the first account does not have corresponding auxiliary information. The account type of the second account is an active type, and the second account has corresponding auxiliary information; namely, the difference of different information types exists between the first account and the second account; furthermore, the auxiliary information of the second account affects the access information of the second account. Even if the auxiliary information of the second account is directly deleted, the difference caused by different information types exists between the first identification information of the first account and the second identification information of the second account.

In one example, the feature reconstruction model is trained by comparing errors to align second identification information and auxiliary information of a second account; and the first reconstruction passing identification characteristic and the second reconstruction passing identification characteristic are obtained by coding processing of a contrast learning encoder, a contrast error is used for aligning the second passing identification information and the auxiliary information, and the contrast error is used for eliminating the difference of different information types between the first account and the second account. The contrast error is used for eliminating interference between the first account and the second account caused by different information types in the feature reconstruction process. The comparison error is used for eliminating the influence caused by different information types between the first account and the second account in the process of clustering the accounts based on the reconstruction characteristics. The comparison error is used for ensuring that the second identification information of the second account is not influenced by the auxiliary information, and ensuring that the first identification information of the first account and the second identification information of the second account belong to the same field.

Fig. 6 shows a flowchart of a training method of a feature reconstruction model according to an exemplary embodiment of the present application. The method may be performed by a computer device. That is, in the embodiment shown in fig. 5, step 510 may be implemented as steps 512 and 514:

step 512: acquiring first identification information of a first account, and acquiring second identification information and auxiliary information of a second account;

the account type of the first account is a first type, the account type of the second account is a second type, and the account in the first type has no corresponding auxiliary information.

Illustratively, the first identification information corresponds to a first account, and the second identification information corresponds to second identification information; the identification information comprises first identification information and second identification information, and the identification information is common information which has a corresponding relationship with both the first account and the second account. The auxiliary information corresponds to a second account, and the identification information is the individual information which only has a corresponding relation with the second account.

In an optional implementation manner, the first identification information and the second identification information are Multi-Hot coded vectors (Multi-Hot vectors); specifically, the method comprises the following steps:

wherein the content of the first and second substances,

the first general-purpose identification information is represented,

representing the mth dimension content in the first identification information, wherein the vector dimension of the first identification information is M dimension; in a similar manner to that described above,

the second identification information is represented by a second identification information,

and the M-dimension content in the second identification information is represented, and the vector dimension of the second identification information is M-dimension.

It is to be understood that in one example the first and second identification information have the same dimension, but the presence of different dimensions is not excluded.

In an alternative implementation, the auxiliary information is a continuous value vector; in one implementation, a continuous-value Vector is also referred to as a Dense Vector (Dense Vector); specifically, the method comprises the following steps:

wherein the content of the first and second substances,

representing auxiliary information, R representing a real number, R ^N The representation auxiliary information is a vector composed of N-dimensional real numbers.

It is to be understood that the dimension N of the auxiliary information may be the same as or different from the dimension M of the first identification information and the second identification information, and the embodiment does not set any limitation thereto.

In an alternative implementation, step 512 may be implemented as: under the condition of independent agreement of an authorized object, acquiring first identification information of a first account, and acquiring second identification information and auxiliary information of a second account;

as will be appreciated by those skilled in the art, in one implementation in the present application, the first and second identification information and the auxiliary information are obtained with the sole consent of the authorized subject to comply with relevant laws and regulations and standards in the relevant country and region. Illustratively, the authorized object is explicitly prompted in at least one of a privacy protocol, a popup prompt and the like, and is acquired after the authorized object is sufficiently authorized; in one implementation, the authorized subject is prompted in each process of acquiring the information, and individual consent of the authorized subject is obtained. In one implementation, the information is presented to the user for use when requesting access to the information from the authorized object. The information is not acquired for exceeding the user's consent scope, and is not uploaded, stored, and disclosed without full authorization from the authorized subject.

Step 514: calling a pass-identification encoder to encode the first pass-identification information to obtain a first pass-identification characteristic, calling the pass-identification encoder to encode the second pass-identification information to obtain a second pass-identification characteristic, and calling an auxiliary encoder to encode the auxiliary information to obtain an auxiliary characteristic;

illustratively, the feature reconstruction model further includes a generic encoder and an auxiliary encoder. Similarly, the generic encoder and the auxiliary encoder include, but are not limited to, at least one of the following networks: convolutional neural networks, cyclic neural networks, long and short term memory networks.

Illustratively, the encoding processes of the general encoder are independent of each other; specifically, the method comprises the following steps:

wherein x is ^(v) Information indicating the first account or the second account, E ^(v) It is shown that the encoder is a digital video encoder,

representing the corresponding parameter of the encoder, z ^(v) Indicating the characteristics of the first account or the second account.

Such as:

wherein E is ⁽¹⁾ A general identification encoder is represented and,

a parameter representing a generic encoder;

the first identification information is represented by a first identification information,

representing a first generic identity feature;

representing a second identifying characteristic.

E ⁽²⁾ It is shown that the auxiliary encoder,

parameters representing an auxiliary encoder;

the auxiliary information is represented by a representation of,

the assist feature is represented.

It should be noted that the encoding processes in this step may be executed simultaneously, or may be executed sequentially in any order; that is, the present embodiment does not make any restrictive provisions on the timing relationship of the encoding processing of the first identification information, the second identification information, and the auxiliary information.

In summary, the method provided in this embodiment expands the construction method of the feature reconstruction model by calling the universal recognition encoder and the auxiliary encoder to perform encoding processing, ensures the effect of clustering accounts, and avoids the problem of difficulty in clustering due to information difference between accounts of different types.

Fig. 7 is a flowchart illustrating a training method of a feature reconstruction model according to an exemplary embodiment of the present application. The method may be performed by a computer device. That is, in the embodiment shown in fig. 6, step 540 may be implemented as step 542, and further includes step 535:

step 535: calling a pass-identification decoder to decode the first pass-identification feature to obtain first reconstruction pass-identification information, calling the pass-identification decoder to decode the second pass-identification feature to obtain second reconstruction pass-identification information, and calling an auxiliary decoder to decode the auxiliary feature to obtain reconstruction auxiliary information;

illustratively, the feature reconstruction model further includes a generic decoder and an auxiliary decoder.

Illustratively, the decoding processes of the generic decoder are independent of each other; specifically, the method comprises the following steps:

wherein, x' ^(v) Reconstruction information representing the first account or the second account, D ^(v) Which represents the decoder, is shown as,

representing the corresponding parameter of the decoder, z ^(v) Indicating the characteristics of the first account or the second account.

Such as:

wherein D is ⁽¹⁾ A generic identification decoder is represented that,

a parameter representing a generic decoder;

the first reconstructed generic identification information is represented,

representing a first generic identity feature;

indicates the second reconfiguration general-purpose information,

representing a second identifying characteristic.

D ⁽²⁾ It is shown that the auxiliary decoder,

parameters representing a secondary decoder;

indicating the side information for the reconstruction is shown,

the assist feature is represented.

It should be noted that the decoding processes in this step may be executed simultaneously, or may be executed sequentially in any order; that is, the present embodiment does not make any restrictive provisions on the timing relationship of the decoding processing performed by the first identifying characteristic, the second identifying characteristic, and the assistant characteristic.

Step 542: training the feature reconstruction model according to the comparison error, the countermeasure error and the reconstruction error to obtain a trained feature reconstruction model;

illustratively, the reconstruction error includes: at least one of an error between the first reconstruction identification information and the first identification information, an error between the second reconstruction identification information and the second identification information, and an error between the reconstruction assistance information and the assistance information.

In an alternative implementation, at least one of the contrast error, the countermeasure error, and the reconstruction error includes a cross-entropy loss.

In an optional implementation manner, the feature reconstruction model is trained according to the sum of the contrast error, the countermeasure error and the reconstruction error, so that the trained feature reconstruction model is obtained.

Such as:

L＝L _rec +L _con +L _adv ；

wherein L represents the sum of contrast error, countermeasure error and reconstruction error, L _con Indicating a contrast error, L _adv Indicates a countermeasure error, L _rec Representing the reconstruction error.

It should be noted that, in this embodiment, step 535 may be performed before, after, or simultaneously with any step in the first branch, where the first branch includes step 522 and step 530; the present embodiment does not provide any limiting specification on the timing relationship between step 535 and the first branch.

In summary, the method provided in this embodiment expands the construction method of the feature reconstruction model by calling the universal recognition decoder and the auxiliary decoder to perform decoding processing, and trains the feature reconstruction model through the reconstruction error, thereby ensuring the effect of clustering accounts and avoiding the problem of difficulty in clustering due to information difference between accounts of different types.

Next, details of the reconstruction error are presented:

in one example, the reconstruction error includes an error between the first reconstruction identification information and the first identification information, an error between the second reconstruction identification information and the second identification information, and an error between the reconstruction assistance information and the assistance information.

It is understood that the reconstruction error includes, but is not limited to, at least one of: cross entropy loss function, 0-1 loss function, dess loss function.

In one example, the reconstruction error includes:

wherein L is _rec Representing a reconstruction error, U representing a sum of the numbers of the first account number and the second account number, M representing dimensions of the first reconstructed identification information and the second reconstructed identification information,

the mth dimension content of the general identification information representing the ith account,

m-dimension content, U, representing reconstructed general identification information of ith account _a Indicating a number representing a second account numberThe amount of the compound (A) is,

auxiliary information indicating the ith second account,

indicating the reconstructed auxiliary information of the ith second account, | | | | | non-woven phosphor ₂ Represents a 2-norm operation, and σ represents a Sigmoid (Sigmoid) function.

Fig. 8 is a flowchart illustrating a training method of a feature reconstruction model according to an exemplary embodiment of the present application. The method may be performed by a computer device. That is, in the embodiment shown in fig. 4, step 530 may be implemented as steps 532 and 534:

step 532: calling a gradient inversion layer to perform inversion processing on the first reconstruction passing identification feature and the second reconstruction passing identification feature to obtain a processed first reconstruction passing identification feature and a processed second reconstruction passing identification feature;

illustratively, the feature reconstruction model includes a gradient inversion layer and an account classifier.

Illustratively, a first reconfiguration common identification feature of the first account and a second reconfiguration common identification feature of the second account are aligned through a Gradient Reverse Layer (GRL) and an account classifier; and training the feature reconstruction model to realize that the second reconstruction general recognition feature does not carry information specific to the field.

Illustratively, the gradient inversion layer includes:

R _λ (χ)＝χ；

wherein I represents a unit matrix, lambda represents a meta-parameter, and chi represents a first reconstruction identification characteristic or a second reconstruction identification characteristic; the gradient inversion layer has no associated parameters, the element parameters are not updated through back propagation, and the gradient inversion layer plays a role in unit transformation in the forward propagation process.

Step 534: calling an account classifier to predict account types of the processed first reconstruction passing identification feature and the processed second reconstruction passing identification feature to obtain a prediction classification result;

illustratively, the account classifier includes, but is not limited to, at least one of: support Vector Machine (SVM), Bayes (Naive Bayes) classifier, decision tree, logic (Logistic) regression.

Illustratively, predicting the classification result includes:

p _active (α|h ⁽¹⁾ )＝softmax(f _adv (α|h ⁽¹⁾ ；φ))；

wherein h is ⁽¹⁾ Representing a reconstruction recognition feature, p _active (·|h ⁽¹⁾ ) The predicted classification result is represented, that is, the probability that the reconstructed generic identity belongs to the target account category is represented, and α represents the target account category, where the target account category is exemplarily a first type corresponding to the first account or a second type corresponding to the second account.

Wherein softmax represents the normalization method function, f _adv Denotes the account classifier and phi denotes the parameters of the account classifier.

Further, the anti-error is introduced:

illustratively, the antagonistic error includes an error between the predicted classification result and the actual classification result; it is understood that the contrast error includes, but is not limited to, at least one of: cross entropy loss function, 0-1 loss function, dess loss function.

In one example, countering the error includes:

wherein L is _adv Representing the countermeasure error, U representing the sum of the numbers of the first and second account numbers, y _i Indicating the actual classification result for the ith account,

indicating the reconstitution of the ith accountIdentifying a feature;

the probability that the account type belongs to the actual classification result of the ith account is expressed under the condition of the reconstruction general identification characteristic of the ith account.

It should be noted that, in one example, the comparison error is used to describe the error between the predicted classification result and the actual classification result; training the feature reconstruction model by resisting errors, and adjusting coding parameters when coding processing is carried out to obtain a first reconstruction passing identification feature and a second reconstruction passing identification feature; and the prediction parameters when predicting the account type are adjusted. Optionally, the first pass identification feature is encoded by calling a contrast learning encoder to obtain a first reconstructed pass identification feature, the second pass identification feature is encoded by calling the contrast learning encoder to obtain a second reconstructed pass identification feature, and the account classifier is called to predict account types of the processed first reconstructed pass identification feature and the processed second reconstructed pass identification feature to obtain a predicted classification result.

Illustratively, the countermeasure error reduces the domain information carried in the first reconstruction general identification feature and the second reconstruction general identification feature, and the feature reconstruction model is trained; the countermeasure error is used for eliminating the domain difference caused by the difference of account types between the first account and the second account; the countermeasure error is used to align the first identification information of the first account with the second identification information of the second account.

In summary, according to the method provided by this embodiment, the gradient inversion layer and the account classifier are called for prediction processing, so that the construction method of the feature reconstruction model is expanded, the trained feature reconstruction model is guided by the gradient inversion layer to be unable to predict to obtain an actual classification result, the effect of clustering accounts is ensured, and the problem of difficulty in clustering due to information difference of accounts of different types is avoided.

Fig. 9 shows a flowchart of a training method of a feature reconstruction model according to an exemplary embodiment of the present application. The method may be performed by a computer device. That is, in the embodiment shown in fig. 4, step 510 may be implemented as step 510a, step 520 may be implemented as step 520a, and step 530 may be implemented as step 530 a:

step 510 a: acquiring a first application characteristic of a first account, and acquiring a second application characteristic and an information characteristic of a second account;

in this embodiment, the first identification information is first application information, the first application information includes installation and uninstallation information of the first account for the at least one application, the second identification information is second application information, the second application information includes installation and uninstallation information of the second account for the at least one application, the auxiliary information is information, and the information includes information browsing information of the second account in the target application.

Illustratively, the at least one application typically comprises the target application, but does not exclude the case where the target application is an application other than the at least one application.

Illustratively, in the installation and uninstallation information of the first account for the at least one application, the at least one application includes at least one of: the first account is a first account that is associated with a first computing device, and the second account is a second account that is associated with a second computing device.

Illustratively, the first application information and the second application information are Multi-Hot encoded vectors (Multi-Hot vectors); the vector dimension of the first application information and the second application information is M-dimension. Wherein the mth dimension content corresponds to installation uninstallation information of the mth application.

Furthermore, when the mth application is installed in the first account, the mth dimension content of the first application information is equal to 1, otherwise, the mth dimension content of the first application information is equal to 0.

Exemplarily, the information is a continuous value vector, and the information behavior of the second account is represented by the continuous value vector; optionally, the information is a concatenation of average pretrained embeddings (embeddings) of clicked and unchecked information.

Illustratively, the information includes, but is not limited to, at least one of the following: articles, card information, video, audio, text information, blog articles, multimedia information. Illustratively, the information affects second application information of the second account, such as: the account for high-frequency multimedia information browsing is provided with more applications of the Xiaozhong.

The first application characteristics comprise characteristic representation of first application information of a first account, the second application characteristics comprise characteristic representation of second application information of a second account, and the information characteristics comprise characteristic representation of information of the second account.

Step 520 a: coding the first application characteristic to obtain a first reconstruction application characteristic, coding the second application characteristic to obtain a second reconstruction application characteristic, and coding the information characteristic to obtain a reconstruction information characteristic;

for example, the encoding processes in this step may be performed simultaneously, or may be performed sequentially in any order; that is, the present embodiment does not make any restrictive provisions on the timing relationship of the encoding processing performed on the first application characteristic, the second application characteristic and the information characteristic.

Step 530 a: predicting account types to which the first reconstruction application characteristic and the second reconstruction application characteristic belong to obtain a prediction classification result;

illustratively, the predicted classification result indicates account types to which the predicted first application generic feature and the second reconstructed application feature belong.

In summary, the method provided in this embodiment ensures the effect of clustering accounts by determining the specific implementation manners of the first general information, the second general information, and the auxiliary information, and improves the information recommendation effect on accounts in the application.

As will be appreciated by those skilled in the art, in another implementation, the first identification information, the second identification information, and the assistance information have at least the following implementations:

in one implementation, the first identification information includes first sales information, the second identification information includes second sales information, and the auxiliary information includes after-sales service information; the first sale information is at least one piece of purchase information of the first account, the second sale information is at least one piece of purchase information of the second account, and the after-sale service information is the after-sale service information corresponding to the at least one piece of purchase information of the second account. Illustratively, the after-sales service information affects second sales information for the second account, such as: the after-sales service information corresponds to the first timestamp, and the after-sales service information affects information after the first timestamp in the second sales information.

Specifically, a first sale characteristic of a first account is obtained, and a second sale characteristic and a service characteristic of a second account are obtained; coding the first sales characteristic to obtain a first reconstructed sales characteristic, coding the second sales characteristic to obtain a second reconstructed sales characteristic, and coding the service characteristic to obtain a reconstructed service characteristic; predicting account types to which the first restructuring sale characteristics and the second restructuring sale characteristics belong to obtain a prediction classification result; and training the feature reconstruction model according to the comparison error and the confrontation error to obtain the trained feature reconstruction model.

In one implementation, the first identification information includes first visitor information, the second identification information includes second visitor information, and the assistance information includes registered member information; the first visitor information is at least one piece of visitor information for accessing the first account, the second visitor information is at least one piece of visitor information for accessing the second account, and the registered member information is at least one piece of visitor information for accessing the second account in the registered member. Illustratively, the registered member information affects second guest information for the second account, such as: for the visitor of the second account, after the registered member exists, the registered member of the second account influences the information of the second visitor of the second account.

Specifically, a first visitor characteristic of a first account is obtained, and a second visitor characteristic and a member characteristic of a second account are obtained; coding the first visitor characteristic to obtain a first reconstructed visitor characteristic, coding the second visitor characteristic to obtain a second reconstructed visitor characteristic, and coding the member characteristic to obtain a reconstructed member characteristic; predicting account types of the first reconstruction visitor feature and the second reconstruction visitor feature to obtain a prediction classification result; and training the feature reconstruction model according to the comparison error and the confrontation error to obtain the trained feature reconstruction model.

FIG. 10 is a flow chart illustrating a method for using a feature reconstruction model provided by an exemplary embodiment of the present application. The method may be performed by a computer device. The method comprises the following steps:

step 610: acquiring the communication characteristic of at least one third account;

illustratively, the passthrough feature comprises a feature representation of the passthrough information of the third account.

Exemplary representations of the feature representations include, but are not limited to: at least one of an eigenvector, an eigenvector matrix, an eigenvalue, or bit information.

The account type of the third account can be a first type or a second type; illustratively, the first type is an inactive type and the second type is an active type.

Step 620: coding the pass identification features based on the feature reconstruction model to obtain reconstructed pass identification features;

illustratively, the feature reconstruction model is obtained by any one of the above-mentioned training methods of the feature reconstruction model.

In an alternative implementation, the feature reconstruction model includes a contrast learning encoder. And calling a contrast learning encoder to encode the pass identification features to obtain reconstructed pass identification features.

Further optionally, the feature reconstruction model further includes a pass identification encoder, and the pass identification feature of the third account is obtained by calling the pass identification encoder to encode the pass identification information of the third account.

Step 630: clustering the reconstructed passing identification characteristics to obtain a clustering result of a third account;

illustratively, the clustering process is implemented by calling a clustering model; wherein the clustering model is constructed using at least one of K-Means (K-Means) clustering, hierarchical clustering, and density clustering.

Illustratively, the clustering result is used to indicate at least one cluster, and at least one third account with the same account characteristics is included in one cluster. Further, accounts in a cluster have the same preference, and such preference may include specific preference content; such as: the account numbers in the first clustering cluster have cartoon preference, and the account numbers in the second clustering cluster have news preference. It is understood that the accounts in the cluster have the same preference or may be abstract preference content, which is only used to indicate that the accounts in the cluster belong to the same cluster.

Optionally, in an implementation manner, the method further includes:

acquiring auxiliary information of a third account in a target time period, and determining account characteristics of a target cluster according to the auxiliary information;

the target cluster includes a cluster indicated by the clustering result of the third account. Illustratively, the account characteristics of the target cluster include an account group representation of the target cluster.

Further optionally, the method further comprises: and calling a sorting model to sort the account characteristics to obtain recommendation information of the account to be recommended.

And the account to be recommended belongs to the target cluster. Adding the account characteristics of the target cluster, namely adding the account group portrait of the target cluster, and obtaining the ranking information of the recommendation information of the account to be recommended through a ranking model.

Optionally, in another implementation, the method further includes:

acquiring high-frequency auxiliary information of a third account exceeding a target threshold in a target time period, and determining recommendation information of an account to be recommended according to the high-frequency auxiliary information;

and the third account and the account to be recommended belong to a target cluster, and the target cluster comprises a cluster indicated by the cluster result of the third account. Illustratively, the high frequency auxiliary information is information that the generation amount or generation frequency exceeds a target threshold in a target time period during the application process of the third account.

In summary, in the method provided in this embodiment, the passing identity feature is reconstructed by the feature reconstruction model, and the reconstructed passing identity feature is subjected to clustering, so that an effect of clustering accounts is ensured, and a problem that clustering is difficult to perform due to information differences of accounts of different types is avoided.

FIG. 11 is a flowchart illustrating an information recommendation method according to an exemplary embodiment of the present application. The method may be performed by a computer device. The method comprises the following steps:

step 660: acquiring application characteristics of a plurality of target application accounts;

the application characteristics include a characteristic representation of application information of the target application account, and the application information includes installation and uninstallation information of the target application account for the at least one application.

Illustratively, the account type of the target application account includes an active type and an inactive type, and in the case that the account type is the active type, the account also corresponds to information, which is information browsing information of the account in the target application. And under the condition that the account type is the inactive type, the account does not have corresponding information.

Illustratively, the target application accounts include at least one fourth account with an inactive account type and at least one fifth account with an active account type,

step 670: coding the application characteristics based on the characteristic reconstruction model to obtain reconstructed application characteristics;

In this step, the application features of the plurality of target application accounts are encoded. In the process of coding based on the feature reconstruction model, even if the account type of the target application account is an active type, only the application features corresponding to the application information of the target application account are coded, and the information features corresponding to the information of the target application account are not coded.

Step 680: clustering the reconstruction application characteristics to obtain a clustering result of the target application account;

illustratively, the clustering process is implemented by calling a clustering model; the clustering result is used for indicating at least one clustering cluster, and at least one account with the same account characteristics is included in one clustering cluster. Illustratively, the clustering result is obtained according to the reconstructed application features, and the clustering result is at least used for indicating the preference of the target application account displayed when the uninstalling application is installed.

It should be noted that the preference, indicated by the clustering result, exhibited when the uninstalling application is installed may be a specific preference, such as: the accounts in the first cluster have the preference of installing the network instant messaging application, and the accounts in the second cluster have the preference of installing the online cooperative office application. The preference indicated by the clustering result may also be an abstract preference, such as merely indicating that accounts in the third cluster of clusters have the same preference, but not directly what preference.

Step 690: determining information recommendation information of a fourth account belonging to the same clustering result as the fifth account based on the information of the fifth account;

illustratively, the information of the fifth account includes information browsing information of the fifth account in the target application. The information recommendation information is recommendation information provided by the fourth account in the target application. The target application is typically, but not exclusively, the application providing the feature reconstruction model, or the host program for the application providing the feature reconstruction model, other applications being fourth account installations and/or logins.

For example, the information recommendation information is presented in at least one of the following manners: multimedia information stream, text information stream, push information, card information, window information.

In one example, the cluster result of the fourth account indicates that a cluster to which the fourth account belongs, and information recommendation information of the fourth account is determined according to information corresponding to a fifth account of which the account type is the active type in the cluster.

The account type of the fifth account is an active type, the fifth account corresponds to information, and the information of the fifth account is information browsing information of the fifth account in the target application.

In summary, in the method provided in this embodiment, the application features are reconstructed by the feature reconstruction model, and the reconstructed application features are clustered, so that the effect of clustering accounts is ensured, and the problem of difficulty in clustering due to information difference between different types of accounts is avoided; the information recommendation information is determined according to the account clustering result, so that the information recommendation information is determined according to the preference of accounts in the same clustering cluster.

Next, the information recommendation information for determining the third account is described in detail, that is, at least the following implementation manner exists in step 690 above:

FIG. 12 is a flowchart illustrating an information recommendation method according to an exemplary embodiment of the present application. The method may be performed by a computer device. That is, in the embodiment shown in FIG. 11, step 690 may be implemented as step 692:

step 692: determining account characteristics of the target cluster according to information of the fifth account in the target time period, and determining information recommendation information of the fourth account according to the account characteristics;

the target cluster comprises a cluster indicated by the cluster result of the fourth account, and the fourth account and the fifth account both belong to the target cluster. Illustratively, the account characteristics of the target cluster include an account group representation of the target cluster. It is understood that the fifth account is generally indicated to belong to the target cluster by the clustering result obtained by the clustering process, but does not exclude the case of determining to belong to the target cluster by other means.

For example, in this embodiment, the account type of the fourth account is an inactive type, the account type of the fifth account is an active type, and the information includes click rate and click rate distribution of the fifth account on information of different types. It can be understood that the information recommendation information of the inactive type account in the target cluster is determined according to the information of the active type account in the target cluster.

In an alternative implementation, the following steps are also included after step 592:

calling a sorting model to sort the account characteristics to obtain sorting information of the information recommendation information;

adding the account characteristics of the target cluster, namely adding the account group portrait of the target cluster, and obtaining the sorting information of the information recommendation information through a sorting model. In one implementation, the account type of the fourth account corresponding to the information recommendation information belongs to an inactive account.

For example, fig. 13 shows an interface diagram of recommendation information provided in an exemplary embodiment of the present application, in the interface diagram 450, according to the account characteristics of the target cluster, the titles are: the first information 452 of the weather forecast for city X is determined to be the first sequential recommendation information, which will be titled: the second information 454 of the Y-cell popularity holiday is determined as the second sequential recommendation information, and will be titled as: the third information 456 of the Z-square weekend promotion is determined as the third sequential recommendation.

In summary, in the method provided in this embodiment, by determining the account characteristics, an account group portrait is constructed for the account of the target cluster, and the information recommendation information of the inactive account in the same cluster is obtained according to the account characteristics, so that the information recommendation effect on the account of the target cluster is improved.

FIG. 14 is a flowchart illustrating an information recommendation method according to an exemplary embodiment of the present application. The method may be performed by a computer device. That is, in the embodiment shown in FIG. 11, step 690 may be implemented as step 694:

step 694: determining information recommendation information of a fourth account according to the high-frequency information that the fifth account exceeds a target threshold within a target time period;

the target cluster includes the cluster indicated by the cluster result of the fourth account. And the fourth account and the fifth account both belong to the target cluster. For example, in this embodiment, the account type of the fourth account is an inactive type, and the account type of the fifth account is an active type.

Illustratively, the high-frequency information is information of which the click rate and/or the click quantity exceeds the click threshold when the fourth account uses the target application program within the target time period. And recalling the high-frequency auxiliary information as a candidate set as information recommendation information of a fourth account.

It can be understood that the information recommendation information of the inactive type account in the target cluster is determined according to the high-frequency information of the active type account in the target cluster.

In summary, the method provided in this embodiment updates the recall candidate set of the inactive type account by acquiring the high-frequency information, so as to ensure the information recommendation effect of the inactive type account in the target cluster.

FIG. 15 is a diagram illustrating an account clustering model according to an exemplary embodiment of the present application.

Training the account clustering model 730 through the first application information 712, the second application information 722 and the information 724 to obtain a trained account clustering model 730 a; illustratively, the account clustering model 730 includes a feature reconstruction model and a clustering model.

The trained account clustering model 730a is called to perform clustering processing on the first application information 712 and the second application information 722 to obtain a clustering result 740 of a first account corresponding to the first application information 712 and a second account corresponding to the second application information 722.

Generating an account group portrait 752 according to the clustering result 740 of the first account and the second account, adding the account group portrait 752 to the ranking model 754, obtaining the sequence information of the recommendation information of the first account belonging to the inactive account in the clustering result 740, and improving the information recommendation effect of the ranking model 754 for the inactive account.

The clustering result 740 of the first account and the second account is added to the recall model 756, and the information that the click rate and the click volume exceed the click threshold in the target time period is used as a candidate set for recall, so that the effect of recalling recommendation for the inactive accounts is improved.

Those skilled in the art will appreciate that the above embodiments can be implemented independently, or can be freely combined to form a new embodiment to implement the method and/or the using method of the feature reconstruction model of the present application.

Fig. 16 is a block diagram illustrating a training apparatus for a feature reconstruction model according to an exemplary embodiment of the present application. The device includes:

an obtaining module 810, configured to obtain a first access characteristic of a first account, and obtain a second access characteristic and an auxiliary characteristic of a second account, where the first access characteristic includes a characteristic representation of first access information of the first account, the second access characteristic includes a characteristic representation of second access information of the second account, and the auxiliary characteristic includes a characteristic representation of auxiliary information of the second account; the account type of the first account is a first type, the account type of the second account is a second type, and the account in the first type has no corresponding auxiliary information;

an encoding module 820, configured to perform encoding processing on the first identification feature to obtain a first reconstructed identification feature, perform encoding processing on the second identification feature to obtain a second reconstructed identification feature, and perform encoding processing on the auxiliary feature to obtain a reconstructed auxiliary feature;

the predicting module 830 is configured to predict account types to which the first reconstruction passing identity feature and the second reconstruction passing identity feature belong, so as to obtain a prediction classification result;

the training module 840 is configured to train the feature reconstruction model according to a comparison error and a countererror, so as to obtain a trained feature reconstruction model, where the comparison error includes an error between the second reconstruction recognition feature and the reconstruction auxiliary feature, the countererror includes an error between the predicted classification result and an actual classification result, and the actual classification result is an account type corresponding to the first recognition feature and the second recognition feature.

In an alternative design of this embodiment, the feature reconstruction model includes a contrast learning encoder;

the encoding module 820 is further configured to:

In an optional design of this embodiment, the feature reconstruction model further includes a general encoder and an auxiliary encoder;

the obtaining module 810 is further configured to:

and calling the identification encoder to encode the first identification information to obtain the first identification feature, calling the identification encoder to encode the second identification information to obtain the second identification feature, and calling the auxiliary encoder to encode the auxiliary information to obtain the auxiliary feature.

In an optional design of this embodiment, the first identification information and the second identification information are multi-hot coded vectors, and the auxiliary information is a continuous value vector.

In an optional design of this embodiment, the feature reconstruction model further includes a general decoder and an auxiliary decoder;

the device further comprises:

a decoding module 850, configured to invoke the pass identity decoder to perform decoding processing on the first pass identity to obtain first reconstructed pass identity information, invoke the pass identity decoder to perform decoding processing on the second pass identity to obtain second reconstructed pass identity information, and invoke the auxiliary decoder to perform decoding processing on the auxiliary feature to obtain reconstructed auxiliary information;

the training module 840 is further configured to:

In an alternative design of this embodiment, at least one of the contrast error, the countermeasure error, and the reconstruction error includes a cross-entropy loss.

In an optional design of this embodiment, the feature reconstruction model includes a gradient inversion layer and an account classifier;

the prediction module 830 is further configured to:

In an optional design of this embodiment, the first identification information is first application information, the first application information includes installation and uninstallation information of the first account for the at least one application, the second identification information is second application information, the second application information includes installation and uninstallation information of the second account for the at least one application, the auxiliary information is information, and the information includes information browsing information of the second account in the target application;

the obtaining module 810 is further configured to:

the encoding module 820 is further configured to:

the prediction module 830 is further configured to:

Fig. 17 is a block diagram illustrating an apparatus for using a feature reconstruction model according to an exemplary embodiment of the present application. The device includes:

an obtaining module 860, configured to obtain a corresponding identification feature of at least one third account, where the corresponding identification feature includes a feature representation of corresponding identification information of the third account;

the encoding module 870 is configured to perform encoding processing on the identification features based on the feature reconstruction model to obtain reconstructed identification features;

a clustering module 880, configured to perform clustering on the reconstructed generic identifier to obtain a clustering result of the third account.

In an optional design of this embodiment, the apparatus further includes:

a determining module 890, configured to obtain auxiliary information of the third account in a target time period, and determine account characteristics of a target cluster according to the auxiliary information, where the target cluster includes a cluster indicated by a cluster result of the third account.

In an optional design of this embodiment, the determining module 890 is further configured to:

In an optional design of this embodiment, the apparatus further includes:

a determining module 890, configured to obtain high-frequency auxiliary information that the third account exceeds a target threshold in a target time period, and determine recommendation information of an account to be recommended according to the high-frequency auxiliary information, where the third account and the account to be recommended belong to a target cluster, and the target cluster includes a cluster indicated by a cluster result of the third account.

FIG. 18 is a block diagram of an information recommendation device according to an exemplary embodiment of the present application. The device includes:

an obtaining module 910, configured to obtain application characteristics of multiple target application accounts, where a target application account includes at least one fourth account whose account type is an inactive type and at least one fifth account whose account type is an active type, where the application characteristics include a characteristic representation of application information of the target application account, and the application information includes installation and uninstallation information of at least one application by the target application account;

the encoding module 920 is configured to perform encoding processing on the application feature based on the feature reconstruction model to obtain a reconstructed application feature;

a clustering module 930, configured to perform clustering processing on the reconstructed application features to obtain a clustering result of the target application account;

the determining module 940 is configured to determine, based on the information of the fifth account, information recommendation information of the fourth account that belongs to the same clustering result as the fifth account, where the information includes information browsing information of the fifth account in a target application.

It should be noted that, when the apparatus provided in the foregoing embodiment implements the functions thereof, only the division of the above functional modules is illustrated, and in practical applications, the above functions may be distributed by different functional modules according to actual needs, that is, the content structure of the device is divided into different functional modules, so as to complete all or part of the functions described above.

With regard to the apparatus in the above-described embodiment, the specific manner in which the respective modules perform operations has been described in detail in the embodiment related to the method; the technical effects achieved by the operations performed by the respective modules are the same as those in the embodiments related to the method, and will not be described in detail here.

An embodiment of the present application further provides a computer device, where the computer device includes: a processor and a memory, the memory having stored therein a computer program; the processor is configured to execute the computer program in the memory to implement the method and/or the using method of the feature reconstruction model provided by the above method embodiments.

Optionally, the computer device is a server. Illustratively, fig. 19 is a block diagram of a server according to an exemplary embodiment of the present application.

In general, the server 2300 includes: a processor 2301 and a memory 2302.

The processor 2301 may include one or more processing cores, such as a 4-core processor, an 8-core processor, and so on. The processor 2301 may be implemented in at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 2301 may also include a main processor and a coprocessor, where the main processor is a processor for Processing data in an awake state, and is also called a Central Processing Unit (CPU); a coprocessor is a low power processor for processing data in a standby state. In some embodiments, the processor 2301 may be integrated with a Graphics Processing Unit (GPU) which is responsible for rendering and drawing the content required to be displayed on the display screen. In some embodiments, the processor 2301 may also include an Artificial Intelligence (AI) processor for processing computing operations related to machine learning.

Memory 2302 may include one or more computer-readable storage media, which may be non-transitory. Memory 2302 may also include high-speed random access memory, as well as non-volatile memory, such as one or more magnetic disk storage devices, flash memory storage devices. In some embodiments, a non-transitory computer-readable storage medium in the memory 2302 is used to store at least one instruction for execution by the processor 2301 to implement the methods and/or methods of use of the feature reconstruction models provided by the method embodiments herein.

In some embodiments, the server 2300 may further optionally include: an input interface 2303 and an output interface 2304. The processor 2301, the memory 2302, the input interface 2303 and the output interface 2304 may be connected by a bus or a signal line. Each peripheral device may be connected to the input interface 2303 and the output interface 2304 via a bus, a signal line, or a circuit board. The Input interface 2303 and the Output interface 2304 can be used for connecting at least one peripheral device related to Input/Output (I/O) to the processor 2301 and the memory 2302. In some embodiments, the processor 2301, memory 2302, and the input and

output interfaces

2303, 2304 are integrated on the same chip or circuit board; in some other embodiments, the processor 2301, the memory 2302, and any one or both of the input interface 2303 and the output interface 2304 can be implemented on separate chips or circuit boards, which is not limited by the embodiments.

Those skilled in the art will appreciate that the above-described illustrated architecture is not meant to be limiting with respect to the server 2300 and may include more or fewer components than those shown, or may combine certain components, or may employ a different arrangement of components.

In an exemplary embodiment, a chip is further provided, which comprises a programmable logic circuit and/or program instructions, and when the chip is run on a computer device, is used for implementing the method for reconstructing a model of features and/or the method for using and/or the method for recommending information according to the above aspects.

In an exemplary embodiment, a computer program product is also provided that includes 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 reads the computer instructions from the computer-readable storage medium and executes the computer instructions to implement the method for reconstructing a model of features and/or the method for using and/or the information recommendation method provided by the above method embodiments.

In an exemplary embodiment, a computer-readable storage medium is further provided, in which a computer program is stored, the computer program being loaded and executed by a processor to implement the method for reconstructing a model of features and/or the method for using and/or the method for recommending information provided by the above-mentioned 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, where the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.

Those skilled in the art will recognize that, in one or more of the examples described above, the functions described in the embodiments of the present application may be implemented in hardware, software, firmware, or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media may be any available media that can be accessed by a general purpose or special purpose computer.

The above description is only exemplary of the present application and should not be taken as limiting, as any modification, equivalent replacement, or improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A training method of a feature reconstruction model is characterized by comprising the following steps:

2. The method of claim 1, wherein the feature reconstruction model comprises a contrast learning encoder;

the encoding the first identification feature to obtain a first reconstructed identification feature, the encoding the second identification feature to obtain a second reconstructed identification feature, and the encoding the assistant feature to obtain a reconstructed assistant feature include:

calling the contrast learning encoder to encode the first pass identification feature to obtain a first reconstruction pass identification feature, calling the contrast learning encoder to encode the second pass identification feature to obtain a second reconstruction pass identification feature, and calling the contrast learning encoder to encode the auxiliary feature to obtain a reconstruction auxiliary feature.

3. The method of claim 2, wherein the feature reconstruction model further comprises a general encoder and an auxiliary encoder;

the acquiring a first identification feature of a first account, and acquiring a second identification feature and an auxiliary feature of a second account includes:

4. The method of claim 3, wherein the feature reconstruction model further comprises a generic decoder and an auxiliary decoder;

the method further comprises the following steps:

calling the pass identification decoder to decode the first pass identification feature to obtain first reconstruction pass identification information, calling the pass identification decoder to decode the second pass identification feature to obtain second reconstruction pass identification information, and calling the auxiliary decoder to decode the auxiliary feature to obtain reconstruction auxiliary information;

the training of the feature reconstruction model according to the contrast error and the confrontation error to obtain the trained feature reconstruction model comprises the following steps:

5. The method according to any one of claims 1 to 4, wherein the feature reconstruction model comprises a gradient inversion layer and an account classifier;

the predicting the account types to which the first reconstruction passing identity feature and the second reconstruction passing identity feature belong to obtain a prediction classification result includes:

6. The method according to claim 1, wherein the first identification information is first application information, the first application information includes installation and uninstallation information of the first account for at least one application, the second identification information is second application information, the second application information includes installation and uninstallation information of the second account for at least one application, the auxiliary information is information, the information includes information browsing information of the second account in a target application;

the predicting the account types to which the first reconstruction passing identity feature and the second reconstruction passing identity feature belong to obtain a predicted classification result includes:

and predicting account types to which the first reconstruction application features and the second reconstruction application features belong to obtain the prediction classification result.

7. A method for using a feature reconstruction model, wherein the feature reconstruction model is trained by the method of any one of claims 1 to 6;

the method comprises the following steps:

8. An information recommendation method, the method comprising:

coding the application features based on a feature reconstruction model to obtain reconstructed application features, wherein the reconstruction model is obtained by training according to the method of any one of claims 1 to 6;

9. An apparatus for training a feature reconstruction model, the apparatus comprising:

10. An apparatus for using a feature reconstruction model, the apparatus comprising:

11. An information recommendation apparatus, comprising:

12. A computer device, characterized in that the computer device comprises: a processor and a memory, wherein at least one program is stored in the memory; the processor is configured to execute the at least one program in the memory to implement the method for training the feature reconstruction model according to any one of claims 1 to 8, or the method for using the feature reconstruction model according to claim 7, or the method for recommending information according to claim 8.

13. A computer-readable storage medium, wherein the computer-readable storage medium stores executable instructions, which are loaded and executed by a processor, to implement the method for training the feature reconstruction model according to any one of claims 1 to 8, or the method for using the feature reconstruction model according to claim 7, or the method for recommending information according to claim 8.

14. A computer program product, characterized in that the computer program product comprises computer instructions stored in a computer-readable storage medium, from which a processor reads and executes the computer instructions to implement the training method of the feature reconstruction model according to any one of the preceding claims 1 to 8, or the use method of the feature reconstruction model according to claim 7, or the information recommendation method according to claim 8.