CN110351318B

CN110351318B - Application recommendation method, terminal and computer storage medium

Info

Publication number: CN110351318B
Application number: CN201810301039.5A
Authority: CN
Inventors: 刘龙坡; 万伟; 陈谦
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2018-04-04
Filing date: 2018-04-04
Publication date: 2022-04-29
Anticipated expiration: 2038-04-04
Also published as: CN110351318A

Abstract

The application discloses a method, a terminal and a computer storage medium for recommending applications, wherein the method comprises the following steps: acquiring application characteristics of an application to be recommended, wherein the application characteristics comprise a first attribute characteristic and a second attribute characteristic; determining an application recommendation network model according to a first attribute feature of an application to be recommended, learning a second attribute feature of the application to be recommended through the application recommendation network model, and determining an application recommendation value or a user recommendation value corresponding to the application to be recommended, wherein the application recommendation network model is obtained by training a sample application feature associated with a first user or by training a sample user feature associated with the first application; and determining the application priority for recommending the application to be recommended to the first user according to the application recommendation value, or determining the user priority for recommending the first application according to the user recommendation value. By adopting the embodiment of the application, the feasibility of the application oriented recommendation can be improved, and the accuracy of the application recommendation can be improved.

Description

Application recommendation method, terminal and computer storage medium

Technical Field

The present application relates to the field of internet technologies, and in particular, to an application recommendation method, a terminal, and a computer storage medium.

Background

With the increasingly powerful functions of terminals such as smart phones and tablet computers, application programs (or applications for short) applied to terminals such as smart phones and tablet computers are increasingly diversified. The application variety diversification brings rich and diverse user experience to the terminal user, and meanwhile, the selection difficulty is added to the terminal user. In order to rapidly and conveniently select applications meeting the requirements or preferences of terminal users from various applications, recommendation services for various applications are developed.

In the prior art, application recommendation is generally performed with higher similarity based on applications already downloaded by a user, or applications that may be of interest are recommended to the user based on the needs or preferences of a user group with mutual experience or interest. However, whether the applications are downloaded by the user or recommended by other applications based on the needs or preferences of the user group, the similarity of the selected applications is high, the occurrence probability of redundant applications is high, the shielding probability of the applications which are not concerned by the user before is high, and further the effectiveness of application recommendation is low and the applicability is poor.

Disclosure of Invention

The embodiment of the application recommendation method, the terminal and the computer storage medium can enhance the user relevance of application recommendation, improve the feasibility of application oriented recommendation, improve the accuracy of application recommendation and have stronger applicability.

In a first aspect, the present application provides a method for application recommendation, including:

acquiring application characteristics of an application to be recommended, wherein the application characteristics of the application to be recommended comprise a first attribute characteristic and a second attribute characteristic, the first attribute characteristic is a first user attribute characteristic, the second attribute characteristic is a second application attribute characteristic, or the first attribute characteristic is a first application attribute characteristic, and the second attribute characteristic is a second user attribute characteristic;

determining an application recommendation network model according to the first attribute characteristics of the application to be recommended, learning second attribute characteristics of the application to be recommended through the application recommendation network model, and determining an application recommendation value corresponding to the application to be recommended or a user recommendation value corresponding to the application to be recommended, wherein the application recommendation network model is obtained by training sample application characteristics associated with a first user, or the application recommendation network model is obtained by training sample user characteristics associated with the first application;

and determining the application priority for recommending the application to be recommended to the first user according to the application recommendation value of the application to be recommended, or determining the user priority for recommending the first application according to the user recommendation value of the application to be recommended.

In a possible implementation manner, the method further includes:

when the application priority of the application to be recommended is greater than or equal to a preset application priority threshold, recommending the application to be recommended to the first user; or

And when the user priority of the application to be recommended is greater than or equal to a preset user priority threshold, determining to recommend the first application to a second user.

In a possible implementation manner, the first attribute feature is a first user attribute feature, and the second attribute feature is a second application attribute feature;

the determining an application recommendation network model according to the first attribute feature of the application to be recommended includes:

matching the first user attribute features with user attribute features associated with each application recommendation network model in an application recommendation network model set, and determining an application recommendation network model associated with the first user corresponding to the first user attribute features from the application recommendation network model set;

the application recommendation network model set further includes other application recommendation network models associated with the user attribute features of other users other than the first user, and the other application recommendation network models are obtained by training sample application features associated with the other users.

In a possible implementation manner, before the obtaining the application feature of the application to be recommended, the method further includes:

acquiring sample data of at least two sample applications for application recommendation training, wherein the sample data of any sample application comprises the first user attribute data and the sample application attribute data;

constructing at least one sample application characteristic pair according to sample data applied by the at least two samples, wherein one sample application characteristic pair comprises a positive sample characteristic and a negative sample characteristic, the positive sample characteristic comprises the first user attribute characteristic and the positive sample application attribute characteristic, and the negative sample characteristic comprises the first user attribute characteristic and the negative sample application attribute characteristic;

and constructing an application recommendation network model according to the at least one sample application characteristic pair.

In a possible implementation manner, the sample application attribute data of each sample application includes activity degree indication information;

the constructing at least one sample application feature pair according to the sample data of the at least two sample applications includes:

pairing the at least two sample applications to obtain at least one sample application pair;

and performing the following operation on any sample application pair i in the at least one sample application pair to obtain at least one sample application characteristic pair:

determining a positive sample application and a negative sample application according to the activity degree indication information of the sample application to the two sample applications of i, wherein the activity degree of the positive sample application is higher than that of the negative sample application;

constructing a positive sample characteristic i according to the first user attribute data and the sample application attribute data of the positive sample application₁And constructing a negative sample characteristic i according to the first user attribute data and the sample application attribute data of the negative sample application₀Obtaining the sample application characteristic pair i corresponding to the sample application pair i₁₀；

Wherein the sample applies the feature pair i₁₀Including the above positive sample feature i₁And the above negative sample characteristic i₀。

In a possible implementation manner, the building an application recommendation network model according to the at least one sample application feature pair includes:

taking the positive sample characteristics and the negative sample characteristics of each sample application characteristic pair in the at least one sample application characteristic pair as input of an application recommendation network model, and learning the positive sample characteristics and the negative sample characteristics of each sample application characteristic pair through the application recommendation network model to obtain the capability of predicting an application recommendation value corresponding to any application characteristic;

and the application recommendation value corresponding to the positive sample feature in any sample application feature pair is larger than the application recommendation value corresponding to the negative sample feature.

In a possible implementation manner, the learning, by the application recommendation network model, the second attribute feature of the application to be recommended, and the determining an application recommendation value corresponding to the application to be recommended includes:

and inputting a second application attribute characteristic of the application to be recommended and the first user attribute characteristic into the application recommendation network model, learning the second application attribute characteristic through the application recommendation network model, and outputting an application recommendation value for recommending the application to be recommended to the first user corresponding to the first user attribute characteristic.

In a possible implementation manner, any sample application corresponds to a first preset recommended value for a positive sample application in the i, and a second preset recommended value for a negative sample application;

applying feature pairs i to any sample₁₀Positive sample characteristic i₁And negative sample characteristics i₀After the application recommendation network model is input, the method further includes:

acquiring positive sample characteristics i output by the application recommendation network model₁Corresponding first application recommendation value, and negative sample characteristic i₀A corresponding second application recommendation value;

calculating a recommendation value loss by combining the difference value of the first preset recommendation value and the second preset recommendation value according to the difference value of the first application recommendation value and the second application recommendation value;

and correcting the application recommendation network model according to the recommendation value loss, and adjusting the prediction precision of the application recommendation network model to the application recommendation value corresponding to any application.

In a possible implementation manner, the first attribute feature is a first application attribute feature, and the second attribute feature is a second user attribute feature;

matching the first application attribute features with application attribute features associated with each application recommendation network model in an application recommendation network model set, and determining an application recommendation network model associated with the first application corresponding to the first application attribute features from the application recommendation network model set;

the application recommendation network model set further includes other application recommendation network models associated with the application attribute features of other applications except the first application, and the other application recommendation network models are obtained by training sample user features associated with the other applications.

acquiring user data of at least two sample users for application recommendation training, wherein the user data of any sample user comprises the first application attribute data and the sample user attribute data;

and constructing at least one sample user characteristic pair according to the user data of the at least two sample users, and constructing an application recommendation network model according to the at least one sample user characteristic pair, wherein one sample user characteristic pair comprises a positive sample characteristic and a negative sample characteristic, the positive sample characteristic comprises the first application attribute characteristic and the positive sample user attribute characteristic, and the negative sample characteristic comprises the first application attribute characteristic and the negative sample user attribute characteristic.

In a possible implementation manner, the sample user attribute data of each sample user includes activity degree indication information;

the constructing at least one sample user feature pair according to the user data of the at least two sample users includes:

pairing the at least two sample users to obtain at least one sample user pair;

and performing the following operation on any sample user pair i in the at least one sample user pair to obtain at least one sample user characteristic pair:

determining a positive sample user and a negative sample user according to the activity degree indication information of the sample user to the two sample users of i, wherein the activity degree of the positive sample user is higher than that of the negative sample user;

constructing a positive sample feature i according to the first application attribute data and the sample user attribute data of the positive sample user₁And constructing an negative sample characteristic i according to the first application attribute data and the sample user attribute data of the negative sample user₀Obtaining a sample user characteristic pair i corresponding to the sample user pair i₁₀；

Wherein, the sample user characteristic pair i₁₀Including the above positive sample feature i₁And the above negative sample characteristic i₀。

In a possible implementation manner, the building an application recommendation network model according to the at least one sample user characteristic pair includes:

taking the positive sample characteristics and the negative sample characteristics of each sample user characteristic pair in the at least one sample user characteristic pair as input of an application recommendation network model, and learning the positive sample characteristics and the negative sample characteristics of each sample user characteristic pair through the application recommendation network model to obtain the capability of predicting a user recommendation value corresponding to any application characteristic;

and the user recommendation value corresponding to the positive sample feature in any sample user feature pair is larger than the user recommendation value corresponding to the negative sample feature.

In a possible implementation manner, the learning, by the application recommendation network model, the second attribute feature of the application to be recommended, and the determining the user recommendation value corresponding to the application to be recommended includes:

and inputting the second user attribute feature of the application to be recommended and the first application attribute feature into the application recommendation network model, learning the second user attribute feature through the application recommendation network model, and outputting a user recommendation value for recommending the first application to the second user corresponding to the second user attribute feature.

In a possible implementation manner, any sample user corresponds to a first preset recommendation value for a positive sample user in the i, and a negative sample user corresponds to a second preset recommendation value;

in any sample user feature pair i₁₀Positive sample characteristic i₁And negative sample characteristics i₀After the application recommendation network model is input, the method further includes:

acquiring positive sample characteristics i output by the application recommendation network model₁Corresponding first user recommendation value, and negative example feature i₀A corresponding second user recommendation value;

calculating a recommended value loss by combining a difference value of the first preset recommended value and the second preset recommended value according to the difference value of the first user recommended value and the second user recommended value;

and correcting the application recommendation network model according to the recommendation value loss, and adjusting the prediction precision of the application recommendation network model on the user recommendation value corresponding to any user.

In a possible implementation manner, the user attribute characteristics of the first user and/or the second user are determined by any user attribute data of user age, user gender, user study, user location area, and user application account;

the application attribute feature of the application to be recommended and/or the sample application attribute feature of the sample application are determined by at least one application attribute data of application identification, application type, activity degree indication information, application resource type and user behavior data.

In a second aspect, the present application provides a terminal, comprising:

the system comprises a feature obtaining unit, a feature obtaining unit and a feature selecting unit, wherein the feature obtaining unit is used for obtaining application features of applications to be recommended, the application features of the applications to be recommended comprise a first attribute feature and a second attribute feature, the first attribute feature is a first user attribute feature, the second attribute feature is a second application attribute feature, or the first attribute feature is a first application attribute feature, and the second attribute feature is a second user attribute feature;

a feature processing unit, configured to determine an application recommendation network model according to the first attribute feature of the application to be recommended acquired by the feature acquisition unit, learn a second attribute feature of the application to be recommended through the application recommendation network model, and determine an application recommendation value corresponding to the application to be recommended or a user recommendation value corresponding to the application to be recommended, where the application recommendation network model is obtained by training a sample application feature associated with a first user, or the application recommendation network model is obtained by training a sample user feature associated with a first application;

and a recommendation predicting unit, configured to determine, according to the application recommendation value determined by the feature processing unit, an application priority for recommending the application to be recommended to the first user, or determine, according to the user recommendation value determined by the feature processing unit, a user priority for recommending the first application.

In a possible implementation manner, the recommendation prediction unit is further configured to:

the feature processing unit is configured to:

matching the first user attribute characteristics acquired by the characteristic acquisition unit with user attribute characteristics associated with each application recommendation network model in an application recommendation network model set, and determining an application recommendation network model associated with the first user corresponding to the first user attribute characteristics from the application recommendation network model set;

In a possible implementation manner, the feature obtaining unit is further configured to:

the feature acquisition unit is configured to:

applying a sample according to the first user attribute data and the positive sampleConstruction of positive sample features i with attribute data₁And constructing a negative sample characteristic i according to the first user attribute data and the sample application attribute data of the negative sample application₀Obtaining the sample application characteristic pair i corresponding to the sample application pair i₁₀；

In a possible implementation manner, the feature obtaining unit is configured to:

In a possible implementation manner, the feature processing unit is configured to:

the feature processing unit is further configured to:

the feature processing unit is configured to:

the feature acquisition unit is configured to:

pairing the at least two sample users to obtain at least one sample user pair;

the feature processing unit is further configured to:

In a possible implementation manner, the user attribute characteristics of the first user and/or the second user are determined by any user attribute data of user age, user gender, user study, user location area and user application account;

In a third aspect, the present application further provides a computer storage medium, where the computer storage medium stores a plurality of instructions, and when the instructions are executed on a terminal, the terminal is caused to perform the method provided by the first aspect and/or any possible implementation manner of the first aspect.

In a fourth aspect, the present application further provides a terminal, including a processor and a memory, where the processor and the memory are connected to each other, where the memory is used to store a computer program that supports the terminal to execute the method provided in the first aspect and/or any one of the possible implementations of the first aspect, where the computer program includes program instructions, and the processor is configured to call the program instructions to execute the method provided in the first aspect and/or any one of the possible implementations of the first aspect.

According to the application recommendation method and device, the user attribute characteristics of the first user and/or the application attribute characteristics of the first application can be used for building the application characteristics of the first application, the application characteristics and/or the user characteristics of the first application are learned through the application recommendation network model to predict the recommendation value when the first application is recommended to the first user and/or the recommendation value when the first application is recommended to the second user in a targeted mode, and therefore targeted application recommendation of the first user and/or targeted user recommendation of the first application can be achieved, and operation is simple. According to the method and the device, the user attribute characteristics of the first user and/or the application attribute characteristics of the first application are integrated into the recommendation process of the first application, the association affinity of the first application recommendation and the first application and/or the first user is enhanced, the probability of recommending non-user-demand or favorite games to the first user can be further reduced, the application recommendation accuracy is improved, meanwhile, the application recommendation redundancy rate can be reduced, and the user stickiness of the terminal is enhanced.

Drawings

In order to more clearly illustrate the technical solution of the present application, the drawings used in the description of the embodiments of the present application will be briefly introduced below.

Fig. 1 is a schematic view of an application scenario of an application recommendation method according to an embodiment of the present application;

fig. 2 is a schematic flow chart of an application recommendation method according to an embodiment of the present application;

fig. 3 is a schematic diagram of another application scenario of the application recommendation method provided in the embodiment of the present application;

fig. 4 is another schematic flowchart of an application recommendation method provided in an embodiment of the present application;

FIG. 5 is a schematic diagram illustrating construction of an application recommendation network model according to an embodiment of the present application;

FIG. 6 is another flowchart illustrating an application recommendation method according to an embodiment of the present application;

FIG. 7 is another flowchart illustrating an application recommendation method according to an embodiment of the present application;

fig. 8 is a schematic structural diagram of a terminal provided in an embodiment of the present application;

fig. 9 is another schematic structural diagram of a terminal according to an embodiment of the present application.

Detailed Description

The technical solutions provided in the embodiments of the present application will be clearly and completely described below with reference to the drawings provided in the embodiments of the present application.

The terminal to which the application recommendation method (which may be referred to as the application recommendation method or method for convenience of description) provided in the embodiment of the present application is applicable includes, but is not limited to, a smart phone, a computer, a tablet computer, a Personal Digital Assistant (PDA), a Mobile Internet Device (MID), a wearable device, and the like. Optionally, the terminal may also be a server corresponding to the smart phone, the computer, the tablet computer, the PDA, the MID, the wearable device, or the like, and may specifically be determined according to an actual application scenario, which is not limited herein. For convenience of description, an execution subject of the application recommendation method provided in the embodiment of the present application will be described with a terminal. Correspondingly, the application recommendation device (or simply, the application recommendation device) provided in the embodiment of the present application includes, but is not limited to, a smart phone, a computer, a tablet computer, a PDA, an MID, a wearable device, and the like. For convenience of description, the application recommendation device and/or the terminal provided in the embodiments of the present application will be described by taking a smart phone (or a mobile phone for short) as an example.

The application recommendation method provided by the embodiment of the application can be applied to recommendation of multiple types of applications, including recommendation of any one type of multiple applications or recommendation of multiple types of multiple applications, and is not limited herein. The above-mentioned various types of applications (for example, mobile phone applications) include, but are not limited to: game-like applications, health-like applications, shopping-like applications, tool-like applications, multimedia-like applications, social-like applications, travel-like applications, educational applications, and the like, without limitation. The above-mentioned applications of the same type may include multiple applications, which are not limited herein. For example, the game applications include, but are not limited to, royal glory, QQ car, fun and fighter, and cool running on heaven. Such health-related applications include, but are not limited to, fitness applications, gourmet applications, and vital sign (e.g., blood pressure) recording applications, among others. The shopping applications include, but are not limited to, movie ticket buying applications, food ordering applications, and living necessities purchasing applications. The tool applications include, but are not limited to, file editing, mail, alarm clock, calendar, photo album, settings, compass, etc. The multimedia applications can include movie players, music players, photography applications, art designing applications, audio recording applications and the like. The social application classes include, but are not limited to WeChat and QQ, among others. Such travel applications include rail 12306, drip travel, and take away travel. The education applications include but are not limited to WeChat reading and QQ reading.

Optionally, the various types of applications in the above examples may exist alone or in a nested manner, and are not limited herein. The separate existence can be understood as applications in an application library (such as an application store), and the applications are classified by the application library. For example, the game-like application may be a game in an application library, such as a multi-style game in an application store. The presence of mutual nesting is understood to mean that an application of a certain type may be one of the application functions in an application. For example, the game-like application and the social-like application may be nested with each other, such that the game-like application becomes one of the application functions of the social-like application. In other words, for example, the WeChat in the social application may be nested with various games such as the King glory in the game class, for example, an application function of the WeChat (WeChat game for short) may be provided with an application function of the game, and the WeChat game may include various games such as the King glory, funny landlord, and the like.

For convenience of description, an application to which the application recommendation method provided in the embodiment of the present application is applied will be described with a game application (game for short) as an example, that is, the application recommendation method provided in the embodiment of the present application may be specifically a game recommendation method. Correspondingly, the application recommendation method provided in the embodiment of the present application may be applicable to application scenarios recommended by multiple games, where the application scenarios recommended by multiple games may include recommendation of any game in an application library, and may also be applicable to recommendation of any game in any application (for convenience of description, application a may be set, for example, wechat), and is not limited herein. An application scenario will be described below with reference to fig. 1, taking the application scenario as game recommendation in application a as an example, where the application recommendation method provided in the embodiment of the present application is applied.

Referring to fig. 1, fig. 1 is a schematic view of an application scenario of an application recommendation method according to an embodiment of the present application. In the embodiment of the application, the terminal (for example, a mobile phone) can monitor the user operation state on the user interface (for example, the interface 1). As shown in fig. 1, it is assumed that terminal applications such as application a, application B, and application C are installed in the terminal. When a terminal user (user for short) clicks an icon of the application A on a user interface of the terminal, the terminal can be triggered to start the user interface (for example, interface 2) of the application A. At this time, the terminal may monitor the user operation instruction on its user interface, and may determine that the application that the user selects to trigger the start is application a according to the click position of the user operation instruction. At this point, the terminal may launch the user interface (e.g., interface 2) of application a. As shown in fig. 1, it is assumed that on the user interface of application a, including multiple switchable interfaces such as chat, contacts, and extended functions, the user can click on different icons on the user interface of application a to switch different functional interfaces of application a. For example, assuming that a user clicks an icon of an extended function, when the terminal detects a user operation instruction on the icon corresponding to the extended function, the user interface of the application a may be switched to display a plurality of extended functions of the application a, such as application function 1, application function 2, …, games, and the like. Further, when the user clicks the operation area corresponding to the game in the extended function of the application a, the terminal may be triggered to switch the user interface of the application a to the display interface (e.g., interface 3) corresponding to the game. On the interface 3, a popularity ranking list of the plurality of games can be displayed, and further the corresponding games can be recommended to the user.

As shown in fig. 1, a recommendation list for recommending the multi-game built in the application a according to a plurality of ranking rules, including a hot game list, a friend hot play list, a new game list, and the like, may be displayed on the interface 3. When the terminal detects a user operation instruction on a display area of any recommendation list, the terminal switches to the recommendation list selected corresponding to the user operation instruction, and a game recommendation list on the recommendation list is output and displayed to the user. For example, when the user clicks the operation area corresponding to "my", the terminal may switch to the personal recommendation list corresponding to "my" for recommending the game to the user. In a specific implementation, how the terminal recommends the personal recommendation list of the game for the user may refer to implementations described in the following embodiments.

Optionally, in the embodiment of the application, when the game is recommended, the multiple games may be ranked according to the user group parameters such as the number of user groups of each game or the attention popularity of the user, and the top ranked multiple games may be recommended to the user. For example, as shown in fig. 1, a hot game board is obtained by ranking a plurality of games according to user group parameters, or a new game board is obtained by ranking a plurality of newly released games according to user group parameters such as a user attention degree, and further 5 games or more in top ranking can be recommended to the user. Or the multiple games are ranked according to friend group parameters such as attention heat of friends of the user to the games and the like to obtain a friend hot-list, and 5 games or more of the list are recommended to the user. The method may be determined according to an actual application scenario, and is not limited herein.

However, when the games with different types are ranked according to the user group parameters such as the attention popularity of the user group to obtain the recommendation list and the top game is recommended to the user, the recommendation list may include the games which the user has paid attention to and played, and may also include other games with higher similarity to the games which the user has played, however, the recommendations are not the games which the user has interested in or needs, so that the recommendation probability of the redundant game is high. Further, for games that are not very popular or new games (i.e., new games) that have not been noticed by the user, it is often difficult to appear on the recommendation chart and thus difficult to find by the user, the exposure of the game is low, and the variety of games selectable by the user is also reduced.

Based on the defects that the redundant probability is high, the recommendation accuracy is low and the like existing in the recommendation list obtained by sequencing the games according to the user group parameters such as the attention popularity of the user group, the embodiment of the application provides a method for realizing the directional user recommendation of the games. According to the game recommendation method provided by the embodiment of the application, the game recommendation model can be constructed by utilizing the user attribute characteristics of a certain user (such as a first user) and the game attribute characteristics of the game played by the first user, and then the game recommendation model can be trained by utilizing the user attribute characteristics of the first user and the game attribute characteristics of the game played by the first user, so that the game recommendation model can determine the recommendation value for recommending each game to the first user aiming at the game characteristics of each game in any one or more games. And determining the recommendation priority of each game when each game is recommended to the target according to the difference of the recommendation values. Wherein, the higher the recommendation value, the higher the game corresponding recommendation priority. According to the method and the device, the user attribute characteristics of the first user, the game attribute characteristics of the game played by the first user and other characteristics can be associated with the recommendation of any other game, so that the directional recommendation of the first user can be realized, the operation is simple, the probability of recommending the game which is not required or favored by the user to the first user can be reduced, the accuracy of game recommendation is improved, the redundancy rate of game recommendation can be reduced, and the user stickiness of the terminal is enhanced.

Optionally, the game recommendation method provided in this embodiment of the application may also use game attribute features of a certain game (which may be referred to as a first game for convenience of description), and game features such as user attribute features of a user group of the first game to construct a game recommendation model, and then may use the game attribute features of the first game and the user attribute features of the user group of the first game to train the game recommendation model so that the game recommendation model may determine, for application features of each game in any one or more games, a recommendation value for directionally recommending the first game to each user in any one or more users. And determining the recommendation priority corresponding to each user when the first game is recommended to each user according to different recommendation values. And the higher the recommendation value is, the higher the recommendation priority corresponding to the user is. According to the game recommending method and device, the game attribute characteristics of the first game, the user attribute characteristics of the first game (namely the second user attribute characteristics) and other application characteristics can be associated with the recommendation of any other user, so that the first game can be directionally recommended to the second user, the operation is simple, the probability of recommending non-user-required or favorite games to the second user can be reduced, the game recommending accuracy is improved, meanwhile, the game recommending redundancy rate can be reduced, and the user stickiness of the terminal is enhanced.

The application recommendation method (e.g., game recommendation method) and apparatus provided in the embodiments of the present application will be specifically described below with reference to fig. 2 to 9.

The first embodiment is as follows:

the application recommendation method provided by the embodiment of the application can be suitable for recommending any application to any user, and for convenience of description, any user can be described by taking a first user as an example, and any application can be described by taking a second application as an example. The method includes the steps that a plurality of applications are assumed to be included in application resources built in a terminal and can be recommended to a first user, wherein the plurality of applications included in the application resources can be combined into an application set to be recommended of the terminal. Each application in the application set to be recommended can correspond to one recommendation priority, the multiple applications can be recommended and sorted according to the recommendation priorities corresponding to the applications, the applications with higher recommendation priorities are sorted in the front, and an application recommendation list for recommending the applications to the first user can be obtained. According to the application recommendation method and device, the recommendation priority of any application in the application set to be recommended can be determined according to the application attribute characteristics of each application in the application set to be recommended and the user attribute characteristics of the first user, and then the display position of the application in the application recommendation list when the application is recommended to the first user can be obtained.

Optionally, in the embodiment of the present application, a recommendation priority corresponding to each user when a certain application (for example, a first application) in the set of applications to be recommended is recommended to each user may be determined according to the application attribute features of each application in the set of applications to be recommended and the user attribute features of the applications to be recommended, so that a recommendation priority corresponding to a second user when the first application is recommended to a certain user (for example, a second user) may be obtained, and it may be determined whether to recommend the first application to the second user. Optionally, if it is determined that the first application is recommended to the second user, a display position of the first application in the application recommendation list may be further determined. For details, reference may be made to the above implementation manner of recommending an application to a first user, and details are not described here.

Referring to fig. 2, fig. 2 is a schematic flow chart of an application recommendation method provided in an embodiment of the present application. As shown in fig. 2, the game recommendation method provided in the embodiment of the present application may include the following steps S201 to S203:

s201, obtaining application characteristics of the application to be recommended.

In some possible embodiments, the application features may include a first attribute feature and a second attribute feature. The first attribute feature may be a first user attribute feature, and the second attribute feature may be a second application attribute feature. Or, optionally, the first attribute feature is a first application attribute feature, and the second attribute feature is a second user attribute feature.

Optionally, the game recommendation method provided in the embodiment of the present application is suitable for recommending an application (e.g., a second application) such as any game to a certain user (e.g., a first user), where the first attribute feature is a first user attribute feature, and the second attribute feature is a second application attribute feature. The first user attribute feature may be specifically a user attribute feature of the first user, and for convenience of description, the user attribute feature of the first user (first user attribute feature for short) may be taken as an example for explanation. The second application attribute feature may be an application attribute feature of the second application, and for convenience of description, the second application attribute feature may be taken as an example for explanation.

Optionally, the game recommendation method provided in the embodiment of the present application is applicable to recommend an application (e.g., a first application) to any user (e.g., a second user), where the first attribute feature is a first application attribute feature and the second attribute feature is a second user attribute feature. The first application attribute feature may be specifically an application attribute feature of the first application, and for convenience of description, the application attribute feature of the first application (first application attribute feature for short) may be taken as an example for explanation. The second user attribute feature may be a user attribute feature of the second user, and for convenience of description, the second user attribute feature may be taken as an example for explanation.

In this embodiment of the present application, in the game recommendation method provided in this embodiment of the present application, an implementation manner in which a certain application (e.g., a first application) is recommended to any user (e.g., a second user) is similar to an implementation manner in which an application such as any game (e.g., a second application) is recommended to a certain user (e.g., a first user), and therefore, an implementation manner in which a second application is recommended to a first user will be described as an example below.

In some possible embodiments, when the terminal recommends an application to the first user, user attribute data of the first user (i.e., first user attribute data) may be first obtained, and a user attribute feature of the first user (i.e., first user attribute feature) may be determined according to the user attribute data of the first user. The user attribute data of the first user may be data for uniquely identifying the first user, and the user attribute data of the first user includes, but is not limited to, an age of the first user, a gender of the first user, a first user study history, a region where the first user is located, a first user application account, and the like. The first user application account may be an application account used by the first user to associate multiple applications to be recommended. For example, the first user application account may be a WeChat account of the first user, and if the multi-payment application to be recommended is a multi-style game in a WeChat game, the first user may associate a part or all of the WeChat game with the WeChat account of the first user. The first user can register a game role in part or all of the WeChat game by using the WeChat account number of the first user, and then the WeChat account number and the game role can be associated.

In some possible embodiments, the terminal may group the user group of the second application according to the user group attribute of the second application, and further may determine, according to the result of the grouping, the first user attribute feature corresponding to the first user by combining the user attribute data of the first user. For example, the terminal may divide the user of the second application into two groups of the male user and the female user according to the user group attribute that both the male user and the female user have in the user group attribute of the second application, and may set a label for each of the male user and the female user, and may further convert the user attribute data into the user attribute feature. For example, assume that the terminals are used to mark a male user and a female user with 1 and 0, respectively, and to mark the gender feature of any user of the second application with a feature parameter of 2 characters. For example, if the user attribute data of the first user is gender data of a male, the terminal may determine that the first user attribute feature of the first user is a male feature according to the user data attribute of the first user, and may further determine that the first user attribute feature of the first user is 10. If the user attribute data of the first user is female gender data, the terminal may determine that the first user attribute feature of the first user is a female feature according to the user data attribute of the first user, and may further determine that the first user attribute feature of the first user is 01. In this way, when the user attribute data of the first user is any one of the data types of the age of the first user, the academic history of the first user, the region where the first user is located, the application account of the first user and the like, and the terminal can also determine the first user attribute characteristics corresponding to the first user attribute data in various expression forms according to the classification mode of the first user attribute data and the conversion mode from the first user attribute data to the first user attribute characteristics. The method may be determined according to an actual application scenario, and is not limited herein.

In some possible embodiments, when the terminal recommends the second application to the first user, the application attribute data of the second application (i.e., the second application attribute data) may be obtained, and the application attribute feature of the second application (i.e., the second application attribute feature) may be determined according to the application attribute data of the second application. The second application attribute data may include various types of application attribute data, including but not limited to an application identifier, an application type, activity indication information, an application resource type, user behavior data, and the like, which is not limited herein. The application identifier may be an application ID, or a number applied in a certain application set (for example, an application set to be recommended), and the like, which is not limited herein. The application types include, but are not limited to, leisure, chess, sports, characters, actions, intelligence, and standalone, wherein the application type of the second application may be one of the above-mentioned application types. The activity indicating information may include an online time length or an online frequency, and the like, wherein the activity indicating information of the second application may include an online time length, an online frequency, and the like for the first user to log in the second application. The application resource type may include a free application, a paid application, or a trial application, and the application resource type of the second application may be one of the multiple application resource types. The user behavior data may be a user operation habit, a user payment record, a user online time period, and the like, and the user behavior data of the second application may be one or more of the user behavior data, which is not limited herein. Optionally, the first user attribute data of the second application may be composed of application attribute data of one of the multiple categories, or may be composed of application attribute data of multiple categories of the multiple categories, and may be specifically based on an actual application scenario, which is not limited herein.

In some feasible embodiments, the terminal may group the second application attribute data of the second application according to the second application attribute data of the second application and the application attribute data of other applications to be recommended in the set of applications to be recommended, and determine a second application attribute feature corresponding to the second application attribute data of the second application according to a grouping result and a conversion manner from the first user attribute data to the first user attribute feature. For example, assuming that the second application is a royal glory in the wechat game, the set of applications to be recommended is an M-money game in the wechat game, where M is a natural number greater than 2. Assuming that the game types of the M game include leisure, chess and cards, competition and intellectual development, the royal players honor the competition among the 4 game types. The terminal can adopt a characteristic parameter with the byte length of 4 characters to convert the game type corresponding to each game into one of the game attribute characteristics corresponding to each game, wherein one character represents one game type and represents yes and no by 1 and 0 respectively. For example, if the game attribute data of the royal glory includes a game type of sports, the terminal may convert the game attribute data of the game of the royal glory type of sports into a game attribute feature 0010 corresponding to the royal glory. 0010 shows whether the type of game that the king glows is leisure (0), chess (0), sports (1), or intellectual development (0). Assuming that the game attribute data of the royal glory includes the game type of leisure and sports, the terminal can convert the game attribute data of the royal glory game of both the leisure type and the sports type into a game attribute feature 1010 corresponding to the glory of the royal. 1010 indicates that the type of game that the joker glows is leisure (1), not chess (0), sports (1), not intellectual (0), etc. The above classification of game types and conversion of game attribute features of royal glory is only an example, and may be determined according to actual application scenarios, and is not limited herein. Similarly, the terminal may determine the second application attribute characteristics corresponding to other second application attribute data of the second application in a similar implementation manner, which is not described herein again.

Optionally, the second application may be an application that the first user has previously focused on, or downloaded, or played, in which case, the user behavior data in the second application attribute feature may be user behavior data recorded and/or stored in the focusing process, or downloading process, or using process of the second application. At this point, the recommendation for the second application may be referred to as a pull-back flow application recommendation. A pull back application recommendation may be understood as an application recommendation intended to call the old user to re-focus on or re-use the second application, in other words as an application recommendation intended to pull back the head-end (user). It is to be understood that applications that the first user has previously focused on or played, but which have not recently focused on or played, are recommended to the first user to elicit a renewed focus on or play of the first user.

Optionally, the second application may be an application that the first user has not paid attention to before, or has downloaded, or has played, and includes a type of application that has been on-line for a long time or has just been on-line or is about to be on-line. At this time, the user behavior data in the second application attribute feature may be null, and the corresponding user behavior feature may be set to 0. Correspondingly, the recommendation of the second application can be referred to as a pull-new application recommendation, and can be understood as an application recommendation mode aiming at pulling a new user. It may be appreciated that applications that the first user has never played before are recommended to the first user to draw attention and/or download play to the first user.

The application recommendation method provided by the embodiment of the application is not only suitable for the application recommendation of the pull-back flow, but also suitable for the application recommendation of the pull-new flow, and is also suitable for the application recommendation of the combination of the two forms or other more forms, and the application recommendation method is not limited, flexible to operate and wide in application range.

In some possible embodiments, after obtaining the first user attribute feature and the second application attribute feature, the terminal may combine the first user attribute feature and the second application attribute feature to construct an application feature corresponding to the second application. The terminal uses the user attribute characteristics of the first user in the construction of the application characteristics of the second application, the recommendation of the second application is closely related to the user attribute characteristics of the first user, the relevance between the recommendation of the second application and the first user is enhanced, the operation is simple, the feasibility of directional recommendation of the second application can be further enhanced, and the applicability is stronger.

Similarly, in the embodiment of the present application, in the implementation manner of recommending the first application to the second user, the application characteristics of the first application may also be constructed according to the implementation manner, where the application characteristics include the first application attribute characteristics and the second user attribute characteristics, so that the recommendation of the first application is closely related to the user attribute characteristics of the second user, and the relevance between the recommendation of the first application and the second user is enhanced. Reference is made in detail to the above-described implementations, which are not limiting herein.

S202, determining an application recommendation network model according to the first attribute characteristics of the application to be recommended, learning the second attribute characteristics through the application recommendation network model, and determining an application recommendation value corresponding to the application to be recommended or a user recommendation value corresponding to the application to be recommended.

In some possible embodiments, after obtaining the application feature of the application to be recommended (for example, the second application), the terminal may match, by using a first user attribute feature included in the application feature of the second application, the application recommendation network model associated with the first user from the plurality of application recommendation network models included in the application recommendation network model set to obtain the application recommendation network model (which may be labeled as the first application recommendation network model for convenience of description). Referring to fig. 3, fig. 3 is a schematic view of another application scenario of the application recommendation method according to the embodiment of the present application. The terminal may store a plurality of application recommendation network models (abbreviated as models, e.g., model 1, model 2, …, model n, etc.) in its database, wherein the plurality of models are used to implement application-oriented recommendations for a plurality of users. For convenience of description, the plurality of application recommendation network models stored in the database may be referred to as an application recommendation network model set, that is, the application recommendation network model set may include the plurality of application recommendation network models such as model 1, model 2, …, model n, and the like. Optionally, the database may also be stored in a server corresponding to the terminal, and may be determined specifically according to an actual application scenario, which is not limited herein. Wherein, each model shown in fig. 3 may adopt the same network architecture, and may be trained by sample application characteristics of sample applications associated with different users to obtain different network parameters. One model can be associated with one user attribute feature, and the user attribute feature associated with each model can be used for marking that the model is applicable to realizing the application recommendation function of the user corresponding to the user attribute feature. As shown in FIG. 3, model 1, model 2, …, model n, etc. may employ the same network architecture, where model 1 may be trained from sample application characteristics of a sample application associated with user 1 to obtain a set of network parameters, such that model 1 with the set of network parameters may be used to implement application-directed recommendations for user 1. Similarly, model 2 may be trained from sample application characteristics of a sample application associated with user 2 to obtain a set of network parameters, such that model 2 with the set of network parameters may be used to implement application-directed recommendations for user 1. Optionally, the model 1, the model 2, the model …, the model n, and the like may also be a same application recommendation network model, and different network parameters may be obtained through training of sample application characteristic data of sample applications associated with different users, so as to implement application-oriented recommendation for different users, which may be determined specifically according to an actual application scenario, and is not limited herein.

Optionally, the terminal may store a plurality of application recommendation network models (abbreviated as models, such as model 1, model 2, …, model n, etc.) in its database, wherein the plurality of models are used for implementing user-directed recommendations for a plurality of applications. The models shown in fig. 3 may adopt the same network architecture, and may be trained by sample user characteristics of sample users associated with different applications to obtain different network parameters. One model can be associated with one application attribute feature, and the application attribute feature associated with each model can be used for marking the application recommendation function of the application which can be suitable for realizing the application corresponding to the application attribute feature. As shown in fig. 3, the same network architecture can be adopted for model 1, model 2, model …, model n, and the like, where model 1 can be trained by sample user characteristics of sample users associated with user 1 to obtain a set of network parameters, so that model 1 with the set of network parameters can be used to implement application-oriented recommendation of application 1 corresponding to the sample user (i.e., recommending application 1 to one or more users, or oriented user recommendation of application 1). Similarly, model 2 may be trained from sample user characteristics of sample users associated with application 2 to obtain a set of network parameters, so that model 2 with the set of network parameters may be used to implement application-oriented recommendation for application 1 (recommending application 2 to one or more users, or oriented user recommendation for application 2). Optionally, the model 1, the model 2, the model …, the model n, and the like may also be a same application recommendation network model, and different network parameters may be obtained through training of sample user feature data of sample users associated with different applications, so as to implement directional user recommendation for different applications, which may be determined specifically according to an actual application scenario, and is not limited herein. The following description will take as an example a model for implementing application-oriented recommendations for different users.

In some possible embodiments, when implementing application-oriented recommendation to the user 1 (which may be set as the first user), a user attribute feature (for example, user attribute feature 1) of the user 1 may be added to an application feature of any application to be recommended to the user 1, and then the model 1 may be obtained by matching the user attribute feature 1 from a plurality of models. Further, the application characteristics of the application to be recommended may be input into the model 1, and further, a recommendation value for recommending the application to be recommended to the user 1 may be output through the model 1. If the application to be recommended is multiple applications, the recommendation value corresponding to each application can be output through the model 1, and then the recommendation priority of each application can be determined according to the recommendation value, so that an application recommendation list is generated. The following description will be given taking as an example a first application to be recommended.

In some possible embodiments, the database shown in fig. 3 may further include application attribute data of each application to be recommended, and user attribute data of a plurality of users. As shown in fig. 3, when a user clicks an application function such as a game, the terminal may detect a user operation instruction on an operation area corresponding to the application function such as the game, and may further generate an application recommendation request, where the application recommendation request may carry user attribute data of the first user. After the terminal generates the application recommendation request, a first application recommendation network model in the multiple application recommendation network models stored in the database can be called according to the first user attribute characteristics corresponding to the user attribute data of the first user. For example, if the first user is user 1, the user attribute features of user 1 may be matched with the user attribute features associated with each application recommendation network model included in the application recommendation network model set, and an application recommendation network model, such as model 1, corresponding to the user attribute features of user 1 is determined from the application recommendation network model set. Since the user attribute feature associated with the model 1 is the user attribute feature of the user 1, the model 1 may also be equivalent to the application recommendation network model associated with the user 1, and is not limited herein.

In some feasible embodiments, after obtaining the application feature of the second application, the terminal may invoke an application recommendation network model associated with the first user according to a first user attribute feature included in the application feature of the second application, input the application feature of the second application into the application recommendation network model, learn, through the application recommendation network model, a second application attribute feature in the application feature of the second application, and predict a recommendation value of the second application based on the application recommendation network model, so as to recommend the second application to the first user. The second application may be one of multiple applications to be recommended, and therefore the second application may form an application recommendation list with other applications to be recommended and output the application recommendation list to a user interface of the terminal, for example, interface 3 in fig. 3. In specific implementation, any one of the application recommendation network models shown in fig. 3 can be obtained by training sample application features of a large number of sample applications, and each application recommendation network model is obtained by training sample applications associated with one user, so that each application recommendation network model is suitable for implementing application-oriented recommendation of one user. For convenience of description, the application recommendation network model corresponding to the first user will be described as an example. The application recommendation network model corresponding to the first user may be obtained by training a large number of sample application features associated with the first user, and each sample application feature in the large number of sample applications includes a sample application attribute feature and the first user attribute feature of the first user. The sample applications may be a plurality of applications used by the first user before the designated time t, and each application may establish an association with the first user through the first user attribute data of the first user when being used as a sample application, and may further be used to train the application recommendation network model for predicting a recommendation value of any application recommended to the first user.

In some possible embodiments, the sample application characteristics of the plurality of sample applications corresponding to the first user may be used as input of an application recommendation network model, and the application characteristics of each sample application are learned by applying the recommendation network model to predict a sample application recommendation value corresponding to the application characteristics of each sample application as a learning task. The recommendation value prediction function of the application recommendation network model can be repeatedly trained through learning of sample application characteristics of multiple sample applications, and therefore the application recommendation network model can have the capability of predicting the application recommendation value corresponding to any input application characteristic. When the terminal predicts the second application recommendation value of the second application by using the trained application recommendation network model, the application characteristics of the second application can be used as the input of the application recommendation network model, and then the application characteristics of the second application can be learned by using the application recommendation network model and the second application recommendation value corresponding to the application characteristics of the second application can be output.

Optionally, in some possible embodiments, the sample user characteristics of the plurality of sample users corresponding to the first application may be used as an input of an application recommendation network model, and the user characteristics of each sample user are learned by predicting a sample user recommendation value corresponding to the user characteristic of each sample user as a learning task through the application recommendation network model. The recommendation value prediction function of the application recommendation network model can be repeatedly trained through the learning of the sample user characteristics of a plurality of sample users, and the application recommendation network model can have the capability of predicting the user recommendation value corresponding to any input user characteristic. When the terminal predicts the second user recommendation value of the second user by using the trained application recommendation network model, the user characteristics of the second user can be used as the input of the application recommendation network model, and then the user characteristics of the second user can be learned by using the application recommendation network model and the second user recommendation value corresponding to the application characteristics of the second user is output. The second user recommendation value may be a recommendation value of a second user when the first application is recommended to the second user, and further, whether to recommend the first application to the second user may be determined according to a comparison of recommendation values of other users except the second user to which the first application is recommended. For details, reference may be made to the above embodiments, which are not described herein again.

S203, determining the application priority for recommending the application to be recommended to the first user according to the application recommendation value of the application to be recommended, or determining the user priority for recommending the first application according to the user recommendation value of the application to be recommended.

In some possible embodiments, when the terminal determines that the second application is recommended to the first user, and after the second application recommendation value recommended by the second application, the terminal may determine the recommendation priority of the second application in the set of applications to be recommended according to the second application recommendation value. The larger the second application recommendation value is, the higher the recommendation priority of the second application in the application set to be recommended is. Correspondingly, the recommendation sequence of the second application in the application recommendation list composed of all the applications to be recommended in the set to be recommended is more advanced. And recommending the second application to the first user when the application priority of the second application is greater than or equal to the preset application priority threshold.

For example, taking an application such as a game as an example, please refer to fig. 1 together, and assuming that when an application is recommended to a first user, a plurality of games such as game 1, game 2, …, game 5, etc. are included in the set of applications to be recommended, the implementation described in the above steps S201 to S203 may implement a personal chart (e.g., "my" in fig. 1) for recommending a game to the first user. Assuming that the second application described in the above steps is game 3, and the recommendation value corresponding to game 3 is higher than the recommendation values of the other games in the set of applications to be recommended, game 3 may be sorted into the first of the personal recommendation list of the first user and the application recommendation list for the first user. The first user can look up the exclusive ranking list of each game pushed by the first user, and further can quickly find the favorite game after looking up popular recommendation information such as the hot ranking list, the friend hot playing list, the new ranking list and the like listed by the user group aiming at each game.

Optionally, in this embodiment of the application, after the terminal determines the second user recommendation value recommended by the second user, the recommendation priority of the second user in the set of users to be recommended may be determined according to the second user recommendation value. The recommendation priority of the second user in the set of users to be recommended is higher when the recommendation value of the second user is larger, and correspondingly, the recommendation sequence of the second user in a user recommendation list formed by all the users to be recommended in the set to be recommended is higher. When the user priority of the second user is larger than or equal to the preset user priority threshold, the first application is determined to be recommended to the second user, so that the first application can be recommended to the second user in a targeted mode, and the effective rate of the user recommending the application of the first application is improved.

In the embodiment of the application, the terminal can use the user attribute feature of the first user in the construction of the application feature of the second application, further can send the application feature constructed by the user attribute feature of the first user and the application attribute feature of the second application into the application recommendation network model, and learns the application feature of the second application through the application recommendation network model to predict the recommendation value of the second application when the second application is recommended to the first user, further can realize the directional recommendation of the first user, and is simple to operate. Further, in the embodiment of the application, the terminal may use the user attribute feature of the second user in the construction of the application feature of the first application, and then may send the application feature constructed by the user attribute feature of the second user and the application attribute feature of the first application into the application recommendation network model, and learn the user feature of the second user through the application recommendation network model to predict the user recommendation value when the first application is recommended to the second user, so that the first application can be directionally recommended to the second user. According to the method and the device, the user attribute characteristics of the first user are integrated into the recommendation process of the second application, and/or the application attribute characteristics of the first application are integrated into the application recommendation process of the first user, so that the association affinity between the application recommendation and the user is enhanced, the probability of recommending non-user-required or favorite games to the user can be reduced, the application recommendation accuracy is improved, the application recommendation redundancy rate can be reduced, and the user stickiness of the terminal is enhanced.

Example two:

in the application recommendation method provided by the embodiment of the application, the terminal realizes recommendation of the application to the first user through the application recommendation network model, and/or realizes directional recommendation of the first application to the second user and other specified user groups through the application recommendation network model, so that the accuracy of application recommendation can be improved, the realization complexity of application recommendation can be greatly reduced, and the efficiency of application recommendation can be improved. The addition of the application recommendation network model greatly reduces the data processing complexity of application recommendation, and simultaneously greatly improves the accuracy and feasibility of application oriented recommendation. The following describes a construction process of the application recommendation network model provided in the embodiment of the present application with reference to fig. 4 and 5. For convenience of description, in the construction process of the application recommendation network model provided in the embodiment of the present application, the first attribute feature is taken as a first user attribute feature, and the second attribute feature is taken as a second application attribute feature. It can be understood that the implementation manner of constructing the application recommendation network model provided in the embodiment of the present application is also applicable to recommendation of a recommended application whose first attribute feature is a first application attribute feature and whose second attribute feature is a second user attribute feature, and is not limited herein.

Fig. 4 is another schematic flow chart of the application recommendation method provided in the embodiment of the present application. In the embodiment of the present application, the construction of the application recommendation network model may include three stages, including stage one: constructing sample characteristics; and a second stage: network model training, and stage three: and (5) testing the network model.

Stage one: sample feature construction may include steps S401-S403 as follows.

S401, sample data of at least two sample applications for applying recommendation training is acquired.

S402, constructing sample characteristics corresponding to the sample data of each sample application in the at least two sample applications.

In a specific implementation, the sample data of any one of the at least two sample applications includes first user attribute data of the first user and sample application attribute data of the sample application, and correspondingly, the sample characteristics of any one of the sample applications includes first user attribute characteristics and sample application attribute characteristics.

In some possible embodiments, when the terminal obtains a sample application for application recommendation training, a plurality of applications that the first user was using before the latest time t may be selected. For example, in selecting a sample game for which the game recommendation is trained, the terminal may select a number of games that the first user was playing before the most recent time t. After the terminal selects the multiple types of applications as the sample applications, the application attribute data of each sample application in the multiple types of sample applications can be recorded, namely the sample application attribute data corresponding to each sample application. It can be understood that, in order to ensure the prediction accuracy of the application recommendation value of any application by the application recommendation network model, in the use phase and the training phase of the application recommendation network model, the construction modes of the application features of the various applications input into the application recommendation network model should be consistent. Therefore, for the record of the sample application attribute data corresponding to each sample application, one or more of multiple application attribute data categories such as an application identifier, an application type, activity indication information, an application resource type, and user behavior data of each sample application can be recorded. The expression form of the application attribute data of each of the multiple application attribute data categories may refer to the corresponding description in step S201, and is not described herein again.

In a possible embodiment, it is assumed that the application attribute data (i.e. sample application attribute data) of each sample application includes multiple items of attribute data such as application identification, application type, and user behavior data. Assuming that the total number of sample applications is M, the above M sample applications may be numbered to obtain M sample applications numbered 1-M. Correspondingly, the application id of each sample application may be a number of each sample application in the M sample applications. Referring to fig. 5, fig. 5 is a schematic diagram illustrating a construction of an application recommendation network model according to an embodiment of the present application. In phase 1 shown in fig. 5, it is assumed that the M sample applications include sample application 2 numbered 2 and sample application 5 numbered 5. The sample application attribute data of the sample application 2 and the sample application 5 each include an application identifier (i.e., the number 2 of the sample application 2 and the number 5 of the sample application), an application type, and user behavior data. The terminal can normalize the application identifiers of the sample applications, namely convert the application identifiers of the sample applications into application identifier characteristics of the same expression mode. For example, the application identification of each sample application is transformed from a dimensional expression to a dimensionless application attribute feature to reduce the learning complexity of the application attribute feature of each sample application by the application recommendation network model. As shown in fig. 5, the terminal may divide the application identifier of each sample application by M to obtain an application identifier characteristic corresponding to the application identifier of each type of identifier. The normalization processing manner shown in fig. 5 is only an example, and may be determined according to an actual application scenario, and is not limited herein.

As shown in fig. 5, the application type may be converted into an application type feature 1001 corresponding to the sample application 2 by the conversion method shown in the first embodiment. Similarly, the application type attribute data included in the sample application attribute data of the sample application 5 may be converted into an application type characteristic 0110 corresponding to the sample application 5. For details, reference may be made to the conversion method from the application attribute data to the application feature data provided in the first embodiment, which is not described herein again.

In some possible embodiments, as shown in fig. 5, user behavior data is also included in the application attribute data of each sample application, such as sample application 2 and sample application 5. The type of the user behavior data corresponding to the second application in the first embodiment is the same, and the user behavior data of each sample application may also include data in various expression forms, such as a registration state, a user operation habit, a user payment record, a user online time length, a user online time period, and the like. For example, the user behavior data included in the application attribute data of the sample application 2 and the sample application 5 shown in fig. 5 are 4 user behavior parameters, such as a registration state, a user operation habit, a user payment record, and a user online time. The registration status includes registered users, unregistered users, and the like, and may be marked with "1" and "0", respectively. The user operation habits comprise a user-defined operation mode and a system-defined operation mode, and are marked by '1' and '0' respectively. The user payment record includes user paid and user not paid and is marked with "1" and "0", respectively. The user online time comprises the online time less than or equal to X hours and the online time greater than X hours, and is marked by 1 and 0 respectively, wherein X is a natural number greater than 0. The terminal determines that the characteristic parameter used for marking the user behavior data is the characteristic parameter of 4 characters, and after the user behavior parameter marked by each character is determined, the user behavior data in the application attribute data of each sample application can be converted into the application attribute characteristics of the sample application according to the user behavior data included in the application attribute data of each sample application. For example, if the user behavior data in the sample application attribute data of the sample application 2 is a registered user, the system defines an operation mode, the user has paid for, and the user online duration is longer than X hours, the terminal may convert the user behavior data of the sample application 2 into the user behavior feature 1011 in the sample application attribute features of the sample application 2. Similarly, the terminal may convert the user behavior data of the sample application 5 into the user behavior feature 0111 in the sample application attribute features of the sample application 5, which is not described herein again.

In a specific implementation, in order to implement the directional application recommendation to the first user, when constructing the sample application attribute features of the sample applications, the terminal needs to add the first user attribute features of the first user to the sample application attribute features of each sample application. As shown in fig. 5, assuming that the user attribute data included in the sample data of the sample application 2 and the sample application 5 both include the attribute data of the first user age, the terminal may segment the ages of the user groups according to information such as age distribution of the user groups of each sample application, so as to group the users. For example, if the user group concerned by each sample application is users whose ages are distributed between 18 and 40 years, the user group may be divided into user groups of 4 age groups, including a group of 18 to 25 years (group 1 may be set for convenience of description), a group of 25 to 30 years (group 2), a group of 30 to 35 years (group 3), and a group of 35 to 40 years (group 4). The terminal may represent the age characteristic of the first user by an age characteristic parameter having a length of 4 characters in bytes, wherein the 4 characters respectively mark the 4 age groups. For example, if the first user's age is 20 years old, it may be determined that the first user's age is included in group 1, and the first user's age may be labeled as "1" in group 1. Wherein, the group 2, the group 3, and the group 4 can be marked with "0" to indicate that the age of the first user is not in the group, and further the first user attribute feature corresponding to the user attribute data (i.e. the age) of the first user is "1000". By analogy, if the user attribute data of the first user is data of other expression forms except for the age, the user attribute data of the first user can be converted into the user attribute feature corresponding to the first user by adopting a similar implementation manner, and the user attribute feature can be specifically determined according to the actual application scenario, and is not described herein again. As shown in fig. 5, since the sample application 2 and the sample application 5 are both sample applications associated with the first user, the sample characteristics corresponding to the sample application 2 and the sample application 5 both include a first user attribute characteristic, i.e., "1000", corresponding to the first user.

In this embodiment of the application, after the terminal determines the first user attribute feature of the first user and the sample application attribute feature of each sample application, the first user attribute feature and the sample application attribute feature of each sample application may be combined to obtain a sample feature corresponding to each sample application. The sample application 2 corresponds to a sample characteristic "10002/M10011011" and the sample application 5 corresponds to a sample characteristic "10005/M01100111" as shown in fig. 5.

And S403, constructing at least one sample application characteristic pair according to the sample data of the at least two sample applications.

In some feasible embodiments, in order to better identify applications in which the first user is more interested from the multiple sample applications, the embodiment of the present application may adopt a pairwise pairing manner between the multiple sample applications to select the applications in which the first user is more interested from the two sample applications, so as to improve the recommendation accuracy of the application recommendation model obtained through sample application training. In a specific implementation, the terminal may pair each two of the M sample applications associated with the first user to obtain a plurality of sample application pairs. Further, positive sample applications and negative sample applications are determined from each sample application pair according to the application attribute data of the respective sample applications. For example, the terminal may determine a positive sample application and a negative sample application from two sample applications of each sample application pair according to the activity level indication information of each sample application, wherein the activity level of the positive sample application is higher than that of the negative sample application. Or, the terminal may determine a positive sample application and a negative sample application from the two sample applications of each sample application pair according to the number of user groups of each sample application, where the number of user groups of the positive sample application is greater than the number of user groups of the negative sample application, and the like. The implementation manner of determining the positive sample application and the negative sample application from the two sample applications of each sample application pair by the terminal according to the sample application attribute data of each sample application can be determined according to an actual application scenario, which is not limited herein. For convenience of description, the activity level indication information will be described as an example of a partition parameter for the positive sample application and the negative sample application.

In some possible embodiments, the terminal may determine the positive sample application and the negative sample application according to the activity indication information of two sample applications in any one sample application pair (for example, the sample application pair i). As shown in fig. 5, assuming that the activity degree indication information is included in the sample application attribute data of each of the sample application 2 and the sample application 5, and it can be determined that the activity degree of the sample application 2 is higher than that of the sample application 5 through the activity degree indication information included in the sample application attribute data of the sample application 2 and the sample application 5, the sample application 2 may be determined as a positive sample application, and the sample application 5 may be determined as a negative sample application. Correspondingly, the sample feature of the positive sample application is the positive sample feature (which may be labeled as positive sample feature i for convenience of description₁) Sample features for negative sample applications are negative sample features (which may be labeled as positive sample features i for ease of description₀) And then obtaining the sample application characteristic pair i corresponding to the sample application pair i₁₀. That is, one sample application feature pair includes one positive sample feature and one negative sample feature, and the corresponding positive sample feature includes the first user attribute feature and the positive sample application attribute feature, and the corresponding negative sample feature includes the first user attribute feature and the negative sample application attribute feature.

In some possible embodiments, after determining that each sample application pair includes the positive sample feature and the negative sample feature, the terminal may construct a sample application feature pair for input training in the application recommendation network model according to the positive sample feature and the negative sample feature, where the construction format may be: (pair < sample _ p, sample _ n >, 1). Where sample _ p represents a positive sample feature, sample _ n represents a negative sample feature, and 1 is a label (label) for marking the positive sample feature. (pair < sample _ p, sample _ n >,1) indicates that for the first user, the interest level of the sample application corresponding to the positive sample feature (the ID of the sample application can be set to p) is greater than that of the sample application corresponding to the negative sample feature (the ID of the sample application can be set to n), so the sample feature corresponding to the sample application p is arranged in front of the sample feature corresponding to the sample application n.

And a second stage: network model training may include steps S404-S405 as follows.

S404, inputting the positive sample characteristic and the negative sample characteristic of each sample application characteristic pair in the at least one sample application characteristic pair into the application recommendation network model.

S405, learning the positive sample characteristics and the negative sample characteristics of the application characteristic pairs of the samples through the application recommendation network model.

In some feasible embodiments, after the terminal constructs at least one sample application feature pair, the positive sample features and the negative sample features in each sample application feature pair can be input into the application recommendation network model, and the positive sample features and the negative sample features of each sample application feature pair are learned through the application recommendation network model, so that the application recommendation network model can obtain the capability of predicting an application recommendation value corresponding to any application feature. The application recommendation value corresponding to the positive sample feature in any sample application feature pair is greater than the application recommendation value corresponding to the negative sample feature, that is, the recommendation priority of the positive sample application is higher than that of the negative sample application. Referring to fig. 5, in the embodiment of the present application, the application recommendation network model may adopt a ranking network model. Because the input of the application recommendation network model during training is the sample application feature pair, that is, the two sample features are included, and the two sample features are input into the same network model, for convenience of description, the training process of the application recommendation network model provided by the embodiment of the application recommendation network model can be explained by adopting a mirror model.

In some possible embodiments, as shown in fig. 5, it is assumed that the application recommendation network model includes two modules, the model structures in the two modules are completely consistent, and the network parameters (e.g., activation functions) of the models are also completely consistent. When the positive sample characteristics and the negative sample characteristics in the sample characteristic pairs are respectively input into the application recommendation network model, the positive sample characteristics and the negative sample characteristics are respectively input into the two modules to be processed to obtain recommendation values (assumed as first application recommendation values) corresponding to the positive sample characteristics and recommendation values (assumed as second application recommendation values) corresponding to the negative sample characteristics. Specifically, the network structure of the application recommendation network model (any module shown in fig. 5) is designed as follows:

the network structure of the application recommendation network model may be designed as a network structure including three fully connected layers, and may also be a network structure of more fully connected layers or a network structure of less fully connected layers in a specific implementation, which is not limited herein. If the number of the full connection layers is too large, the data processing complexity may be too high, and then the prediction result of the network model handles data processing problems such as overfitting. If the number of the full connection layers is too small, the accuracy of data processing may be insufficient, and further, data processing problems such as under-fitting and the like may occur in the prediction result of the network model. Therefore, the number of fully connected layers may be determined according to the application requirements, such as the size of data processing amount in the actual application requirements or the prediction accuracy of the network model, and is not limited herein.

For convenience of description, the following description will be made by taking four fully-connected layers as an example, as shown in fig. 5:

the first layer is a fully-connected layer, which includes N neurons (assuming that the number of neurons is 256), and the activation function of each neuron is a linear rectification function (Relu). The fully-connected layer of the first layer can learn the input sample characteristics (positive sample characteristics or negative sample characteristics) through each included neuron, and then can output the sample characteristics obtained after learning to the neuron of the next layer for processing.

The second layer is also a fully connected layer, which includes N/2 neurons (i.e., 128 neurons), and the activation function of each neuron is relu. In the network structure of the model, the next fully-connected layer learns the input sample features and reduces the dimension of the output sample features of the previous layer so as to reduce redundant features. It can be understood that the dimension reduction of the output feature of the previous full connection layer by the next full connection layer may be understood as that the high-dimensional feature output by the previous full connection layer is converted into the low-dimensional feature by means of feature mapping, etc., that is, some redundant features in the high-dimensional feature are removed to obtain a feature with less data amount, so that the data processing complexity can be reduced while the necessary features are retained. Similarly, the second fully-connected layer may learn the sample features (positive sample features or negative sample features) output by the first fully-connected layer through each neuron included in the second fully-connected layer, and then may output the sample features obtained after learning to the next neuron for processing.

The third layer is also a fully connected layer, which includes N/4 neurons (i.e., the number of neurons is 64), and the activation function of each neuron is relu. The last layer is a fully connected layer of 1 neuron, and the activation function is sigmoid. Wherein, the last full connection layer adopts a sigmoid function, and the output of the model can be limited to be between [0,1] as the fraction (i.e. the recommended value) of the model output.

In some possible implementations, the recommendation value corresponding to the positive sample feature, such as the first application recommendation value, may be output after the application recommendation network model learns the positive sample feature in the input sample feature pairs. Similarly, the recommended value corresponding to the negative sample feature, for example, the second recommended application value, may be output after the application recommended network model learns the negative sample feature. The application recommendation network model can also perform data processing on the first application recommendation value and the second application recommendation value, and can modify the output result by combining the label carried in the sample feature pair so as to continuously strengthen the precision of the predicted recommendation value of the application recommendation network model and improve the accuracy of the output recommendation value of the application recommendation network model. In this embodiment of the application, the recommended value corresponding to the positive sample feature may be set to 1, and the recommended value corresponding to the negative sample feature is also set to 0, so that the label carried in the input of the sample feature may be used to mark the amount of the preset recommended value of the positive sample feature, and may also be used to mark the difference between the recommended values corresponding to the positive sample feature and the negative sample feature. The method may be determined according to an actual application scenario, and is not limited herein.

In some possible embodiments, during the training process of applying the recommended network model, the terminal may set that, in any sample application, a positive sample application corresponds to a first preset recommended value (e.g., 1), and a negative sample application corresponds to a second preset recommended value (e.g., 0). After the positive sample characteristics and the negative sample characteristics of any sample application characteristic pair are input into the application recommendation network model, a first application recommendation value corresponding to the positive sample characteristics and a second application recommendation value corresponding to the negative sample characteristics output by the application recommendation network model can be obtained. The terminal can calculate the loss of the recommended value by combining the difference value of the first preset recommended value and the second preset recommended value according to the difference value of the first application recommended value and the second application recommended value. For example, the terminal may perform difference calculation on the first application recommended value and the second application recommended value to obtain a difference value between the two application recommended values, may further perform sigmoid transformation on the difference value to scale to [0,1], and perform loss calculation on an output of the sigmoid transformation and a tag in the sample feature pair to obtain a recommended value loss. The recommendation value loss can be fed back to the learning process of the application recommendation network model on other sample characteristics, and the recommendation value parameter can be used for optimizing an activation function of the application recommendation network model to correct the prediction accuracy of the application recommendation value corresponding to the other sample characteristics by the application recommendation network model, so that the output recommendation value of the application recommendation network model on the positive sample characteristics in the sample characteristic pair is as close to 1 as possible. That is, the application recommendation value (first application recommendation value) output after learning the positive sample features by the application recommendation network model is made as large as possible for each sample feature pair than the application recommendation value (second application recommendation value) output after learning the negative sample features. By repeating the steps, the prediction accuracy of the application recommendation network model can be continuously corrected, and the application recommendation network model can further have the capability of predicting the recommendation value corresponding to any application characteristic.

And a third stage: the network model test may include the following step S406.

S406, obtaining test characteristics for the application recommendation test and inputting the application recommendation network model.

In some feasible embodiments, after the terminal trains and obtains the application recommendation network model, the application attribute data of the test application of the application recommendation test can be obtained, and then the test characteristics can be obtained according to the application data processing of the test application. The test application for the application recommendation test may be one or more applications used by the first user after time t, which is not limited herein. In specific implementation, after the terminal obtains any test application, the application attribute characteristics corresponding to the test application can be obtained by processing according to the application attribute data of the test application. The conversion manner from the application attribute data of the test application to the application attribute feature of the test application may refer to an implementation manner corresponding to the sample application, and is not described herein again.

Referring to fig. 5, after the terminal processes the application attribute data of the test application to obtain the application attribute feature corresponding to the test application, the terminal may construct, in combination with the first user attribute feature of the first user, an application feature corresponding to the test application, such as "10003/M10110101" corresponding to the test application 3 shown in fig. 5. After obtaining the application characteristics corresponding to the test application 3, the terminal may input the application characteristics to the application recommendation network model, so as to output a recommendation value corresponding to the application characteristics through the application recommendation network model. For convenience of description, referring to fig. 5, the application feature corresponding to the test application 3 may be input into any one of the two modules (essentially, the same module) of the application recommendation network model shown in fig. 5, and after the corresponding recommendation value is output through the application recommendation network model, the terminal may determine the recommendation value output by the application recommendation network model according to the user interest level of the test application 3. And if the judged recommended value output by the application recommendation network model meets the requirement of the actual application scene, determining that the prediction precision of the application recommendation network model can meet the requirement of the actual application scene, wherein the application recommendation network model can be used for the application recommendation use scene of any application. If the recommended value output by the judged application recommendation network model does not meet the requirements of the actual application scene, more sample applications need to be obtained again to carry out more training on the application recommendation network model until the application recommendation network model with higher prediction accuracy is obtained.

In the embodiment of the application, the terminal can train the application recommendation network model through a large number of sample applications to construct a recommendation value with the capability of predicting and recommending any application to the first user in a targeted mode. The recommendation of the application to the first user is realized by applying the recommendation network model, so that the accuracy of application recommendation can be improved, the implementation complexity of application recommendation can be greatly reduced, the efficiency of application recommendation is improved, and the applicability is stronger.

Example three:

referring to fig. 6, fig. 6 is another schematic flowchart of an application recommendation method provided in an embodiment of the present application. In the embodiment of the application, after the terminal trains and obtains the application recommendation network model, the application recommendation network model may be used for application feature learning of any application to be recommended (for example, a second application) and outputs a recommendation value corresponding to the application to be recommended. The recommendation process for any application to be recommended may include the following steps S601-S604:

s601, obtaining application characteristics of the application to be recommended.

S602, inputting the application characteristics of the application to be recommended into the application recommendation network model.

S603, determining an application recommendation value corresponding to the application characteristic through an application recommendation network model.

S604, determining the recommendation priority of the application to be recommended in the application set to be recommended according to the application recommendation value, and outputting an application recommendation list of the first user.

In a specific implementation, the implementation manners performed by each step in steps S601 to S604 may refer to the implementation manners described in each step in the first embodiment, and are not described herein again.

That is, in some possible embodiments, when recommending an application, the application that is interested in the first user is predicted by using the previously trained application recommendation network model, an output score (i.e., a corresponding application recommendation value) of the application recommendation network model represents a degree of interest of the first user in the application, and thus the application that is interested in the first user can be recommended according to the application recommendation value.

In some possible embodiments, when the application recommendation is performed, for pulling the new recommendation, a new application recommendation list for recommending to the first user may be obtained first. Because each application in the new application recommendation list has no user behavior data, the user behavior characteristics in the application attribute characteristics corresponding to each application can be zeroed when the application characteristics recommended by the application to be recommended are constructed. And constructing application characteristics of the applications to be recommended according to the application attribute characteristics of the applications to be recommended and the first user attribute characteristics, and inputting the application characteristics into an application recommendation network model to obtain a recommendation value of each application to be recommended by a user. Further, the terminal can determine the recommendation priority of each application to be recommended according to the recommendation value corresponding to each application to be recommended from high to low, wherein the higher the recommendation value is, the higher the corresponding recommendation priority is. The terminal may rank, according to a recommendation value corresponding to each application to be recommended, each application to be recommended to obtain an application recommendation list for the first user, such as a first user-specific ranking list (i.e., "my") shown in fig. 1.

In some possible embodiments, for the pull-back flow recommendation, user behavior data that is concerned by or used by the user before exists in each application to be recommended, so that the application feature construction corresponding to each application to be recommended may be performed according to the feature construction method described in the foregoing embodiment one and/or embodiment two, and details are not repeated here. Similarly, after the application characteristics of each application to be recommended are obtained through construction, the application characteristics of each application to be recommended can be input into the application recommendation network model, the application recommendation value corresponding to each application to be recommended is determined through the application recommendation network model, the recommendation priority of each application to be recommended can be further determined, and the application recommendation list of the first user is output.

In the embodiment of the application, the terminal can use the user attribute feature of the first user in the construction of the application feature of the second application, further can send the application feature constructed by the user attribute feature of the first user and the application attribute feature of the second application into the application recommendation network model, and learns the application feature of the second application through the application recommendation network model to predict the recommendation value when the second application is recommended to the first user, so that the directional recommendation of the first user can be realized, and the operation is simple. According to the method and the device, the user attribute characteristics of the first user are integrated into the recommendation process of the second application, the association affinity between the application recommendation and the user is enhanced, the probability of recommending non-user-required or favorite games to the user can be reduced, the application recommendation accuracy is improved, the application recommendation redundancy rate can be reduced, and the user stickiness of the terminal is enhanced.

Example four:

it can be understood that the implementation manner of constructing the application recommendation network model provided in the embodiment of the present application is also applicable to recommendation of a recommended application whose first attribute feature is a first application attribute feature and whose second attribute feature is a second user attribute feature, and is not limited herein.

Fig. 7 is another schematic flow chart of the application recommendation method according to the embodiment of the present application. In the embodiment of the present application, the construction of the application recommendation network model may also include three stages, including stage one: constructing sample characteristics; and a second stage: network model training, and stage three: and (5) testing the network model.

Stage one: the sample feature construction may include the following steps S701-S703.

S701, obtaining user data of at least two sample users for applying recommendation training.

In some possible embodiments, the user data of any one of the sample users includes the first application attribute data and the sample user attribute data.

And S702, constructing sample characteristics corresponding to the sample data of each of the at least two sample users.

S703, constructing at least one sample user feature pair according to the user data of the at least two sample users.

In some possible embodiments, a pair of sample user attributes includes a positive sample attribute and a negative sample attribute, where the positive sample attribute includes the first application attribute and the positive sample user attribute, and the negative sample attribute includes the first application attribute and the negative sample user attribute.

In some possible embodiments, the sample user attribute data of each sample user includes activity level indication information;

pairing the at least two sample users to obtain at least one sample user pair;

In a specific implementation, the implementation manner for constructing the sample feature of the first application corresponding to the user from the user data of the at least two sample users may refer to the implementation manner for constructing the sample feature of the first application corresponding to the first user from the application data of the at least two sample applications in the implementation manners provided in steps S401 to S403 in the embodiment shown in fig. 4, and details are not repeated here.

And a second stage: network model training may include the following steps S704-S705.

And S704, taking the positive sample characteristic and the negative sample characteristic of each sample user characteristic pair in the at least one sample user characteristic pair as input of the application recommendation network model.

S705, learning the positive sample characteristics and the negative sample characteristics of the sample user characteristic pairs through the application recommendation network model.

In some possible embodiments, the implementation of training the network model by using the at least one sample user feature may refer to the implementation provided in steps S404 and S405 in the embodiment shown in fig. 4, and will not be described herein again.

And a third stage: the network model test may include the following step S706.

S706, obtaining the test characteristics for the application recommendation test and inputting the application recommendation network model.

In some possible embodiments, the implementation of the training of the network model may refer to the implementation provided in step S46 in the embodiment shown in fig. 4, and will not be described herein again.

In some feasible embodiments, when the terminal learns the second attribute feature of the application to be recommended through the application recommendation network model and determines the user recommendation value corresponding to the application to be recommended, the terminal may further input the second user attribute feature of the application to be recommended and the first application attribute feature into the application recommendation network model, learn the second user attribute feature through the application recommendation network model, and output the user recommendation value recommending the first application to the second user corresponding to the second user attribute feature.

In one possibilityIn the implementation manner of (3), the any sample user corresponds to a first preset recommendation value for the positive sample user in the i, and the negative sample user corresponds to a second preset recommendation value. The terminal can couple any sample user characteristic i₁₀Positive sample characteristic i₁And negative sample characteristics i₀Inputting the application recommendation network model, and obtaining a positive sample characteristic i output by the application recommendation network model₁Corresponding first user recommendation value, and negative example feature i₀A corresponding second user recommendation value. The terminal can calculate recommendation value loss according to the difference value between the first user recommendation value and the second user recommendation value and by combining the difference value between the first preset recommendation value and the second preset recommendation value, revise the application recommendation network model according to the recommendation value loss, and adjust the prediction accuracy of the application recommendation network model on the user recommendation value corresponding to any user.

In some possible embodiments, when the terminal performs directional recommendation on the first application to the second user, the user group interested in the first application is predicted by using the previously trained application recommendation network model, an output score (i.e., a corresponding user recommendation value) of the application recommendation network model represents a degree of interest of the second user interested in the first application, and thus the terminal may perform directional recommendation on the first application to the second user according to the user recommendation value.

In the embodiment of the application, the terminal can use the application attribute feature of the first application in the construction of the application feature corresponding to the second user, and then can send the application feature constructed by the application attribute feature of the first application and the user attribute feature of the second user into the application recommendation network model, and learn the user feature of the second user through the application recommendation network model to predict the recommendation value corresponding to the second user when the first application is recommended to the second user, so that the directional recommendation of the first application to the second user can be realized.

Based on the above description of the method embodiment of application recommendation, the present application embodiment also discloses an application recommendation apparatus (which may be simply referred to as an application recommendation apparatus for convenience of description), where the application recommendation apparatus may be a computer program (including program codes) running in a terminal, and the application recommendation apparatus may be applied in the application recommendation methods of the embodiments shown in fig. 2 to fig. 7 for executing steps in the application recommendation methods. For convenience of description, the following description will be given taking the terminal as an example. Referring to fig. 8, fig. 8 is a schematic structural diagram of a terminal according to an embodiment of the present application. In the embodiment of the application, the terminal can operate the following units:

the feature obtaining unit 81 is configured to obtain an application feature of the application to be recommended.

The application features of the application to be recommended include a first attribute feature and a second attribute feature, where the first attribute feature is a first user attribute feature and the second attribute feature is a second application attribute feature, or the first attribute feature is a first application attribute feature and the second attribute feature is a second user attribute feature.

A feature processing unit 82, configured to determine an application recommendation network model according to the first attribute feature of the application to be recommended acquired by the feature acquiring unit 81, learn a second attribute feature of the application to be recommended through the application recommendation network model, and determine an application recommendation value corresponding to the application to be recommended or a user recommendation value corresponding to the application to be recommended.

The application recommendation network model is obtained by training sample application characteristics associated with a first user, or the application recommendation network model is obtained by training sample user characteristics associated with the first user.

A recommendation predicting unit 83, configured to determine an application priority for recommending the application to be recommended to the first user according to the application recommendation value determined by the feature processing unit 82, or determine a user priority for recommending the first application according to the user recommendation value determined by the feature processing unit.

In a possible implementation manner, the recommendation prediction unit 83 is further configured to:

the feature processing unit 82 is configured to:

matching the first user attribute characteristics acquired by the characteristic acquisition unit 81 with user attribute characteristics associated with each application recommendation network model included in an application recommendation network model set, and determining an application recommendation network model associated with the first user corresponding to the first user attribute characteristics from the application recommendation network model set;

In a possible implementation manner, the feature obtaining unit 81 is further configured to:

In a feasible implementation manner, the sample application attribute data of each sample application includes activity degree indication information;

the feature acquiring unit 81 is configured to:

In a possible implementation manner, the feature obtaining unit 81 is configured to:

In a possible implementation, the feature processing unit 82 is configured to:

In a feasible implementation manner, any sample application corresponds to a first preset recommended value for a positive sample application in the i, and a second preset recommended value for a negative sample application;

the feature acquiring unit 81 is further configured to:

the feature processing unit 81 is configured to:

In a feasible implementation manner, the sample user attribute data of each sample user includes activity degree indication information;

the feature acquiring unit 81 is configured to:

pairing the at least two sample users to obtain at least one sample user pair;

constructing a positive sample feature i according to the first application attribute data and the sample user attribute data of the positive sample user₁And constructing a negative sample according to the first application attribute data and the sample user attribute data of the negative sample userCharacteristic i₀Obtaining a sample user characteristic pair i corresponding to the sample user pair i₁₀；

In a possible implementation, the feature processing unit 82 is configured to:

In a feasible implementation manner, any sample user corresponds to a first preset recommendation value for a positive sample user in the i, and a negative sample user corresponds to a second preset recommendation value;

the feature processing unit 82 is further configured to:

According to the first embodiment of the present application, the implementation manner described in steps S201 to S203 in the application recommendation method shown in fig. 2 may be performed by each unit of the terminal shown in fig. 8. For example, the implementation described in steps S201, S202, and S203 in the application recommendation method shown in fig. 2 described above may be performed by the feature acquisition unit 81, the feature processing unit 82, and the recommendation prediction unit 83 in the terminal shown in fig. 8, respectively. The implementation manners executed by the feature obtaining unit 81, the feature processing unit 82, and the recommendation predicting unit 83 may refer to the implementation manners provided in the steps in the first embodiment, and are not described herein again.

According to the second embodiment of the present application, the implementation described in steps S401 to S406 in the application recommendation method shown in fig. 4 is also performed by the feature acquisition unit 81 and the feature processing unit 82 in the terminal shown in fig. 8. The implementation manners executed by the feature obtaining unit 81 and the feature processing unit 82 may refer to the implementation manners described in each step in the second embodiment, and are not described herein again.

According to the third embodiment of the present application, the implementation manner described in steps S601 to S604 in the application recommendation method shown in fig. 6 can be performed by each unit in the terminal shown in fig. 8. For example, the implementations described in steps S601 to S602, S603, and S604 in the application recommendation method shown in fig. 6 described above may be performed by the feature acquisition unit 81, the feature processing unit 82, and the recommendation prediction unit 83 in the terminal shown in fig. 8, respectively. The implementation manners executed by the feature obtaining unit 81, the feature processing unit 82, and the recommendation predicting unit 83 may refer to the implementation manners provided in the steps in the third embodiment, which are not described herein again.

According to the fourth embodiment of the present application, the implementation described in steps S701 to S706 in the application recommendation method shown in fig. 7 is also performed by the feature acquisition unit 81 and the feature processing unit 82 in the terminal shown in fig. 8. The implementation manners executed by the feature obtaining unit 81 and the feature processing unit 82 may refer to the implementation manners described in each step in the fourth embodiment, and are not described herein again.

In this embodiment of the present application, the units in the terminal shown in fig. 8 may be respectively or entirely combined into one or several other units to form the terminal, or some unit(s) may be further split into multiple units with smaller functions to form the terminal, which may implement the same operation without affecting implementation of technical effects of the embodiment of the present application. The units are divided based on logic functions, and in practical application, the functions of one unit can be realized by a plurality of units, or the functions of a plurality of units can be realized by one unit. In other possible implementations of the present application, the terminal may also include other units, and in practical applications, the functions may also be implemented by being assisted by other units, and may be implemented by cooperation of multiple units, which is not limited herein.

In the embodiment of the application, the terminal can use the user attribute feature of the first user in the construction of the application feature of the second application, further can send the application feature constructed by the user attribute feature of the first user and the application attribute feature of the second application into the application recommendation network model, and learns the application feature of the second application through the application recommendation network model to predict the recommendation value when the second application is recommended to the first user, so that the directional recommendation of the first user can be realized, and the operation is simple. Further, in the embodiment of the application, the terminal may use the user attribute feature of the second user in the construction of the application feature of the first application, and then may send the application feature constructed by the user attribute feature of the second user and the application attribute feature of the first application into the application recommendation network model, and learn the user feature of the second user through the application recommendation network model to predict the user recommendation value when the first application is recommended to the second user, so that the first application can be directionally recommended to the second user. According to the method and the device, the user attribute characteristics of the first user are integrated into the recommendation process of the second application, and/or the application attribute characteristics of the first application are integrated into the application recommendation process of the first user, so that the association affinity between the application recommendation and the user is enhanced, the probability of recommending non-user-required or favorite games to the user can be reduced, the application recommendation accuracy is improved, the application recommendation redundancy rate can be reduced, and the user stickiness of the terminal is enhanced.

Based on the application recommendation method shown in the foregoing embodiment, an embodiment of the present application further provides a terminal, where the terminal may be applied to the application recommendation method shown in the embodiments shown in fig. 2 to 7, so as to execute the steps in the application recommendation method.

Referring to fig. 9, fig. 9 is another schematic structural diagram of a terminal provided in the embodiment of the present application. In the embodiment of the present application, the terminal may include a processor 91, a computer storage medium (or memory) 92, and a communication interface 93. The computer storage medium (or memory) 92 is used to store a computer program supporting the application recommendation method provided in each of the above embodiments, where the computer program may be one or more program instructions (or instructions for short, such as instruction 1, instruction 2, and instruction …, instruction N). The processor 91, the computer storage medium (or memory) 92 and the communication interface 93 may be connected by a bus 94 or other means, and fig. 9 shown in the embodiment of the present application is exemplified by being connected by the bus 94.

The communication interface is a medium for realizing interaction and information exchange between the terminal and external equipment. Processors (e.g., Central Processing Units (CPUs)) are a computing core and a control core of the terminal, and are adapted to implement one or more instructions, and in particular, are adapted to load and execute one or more instructions so as to implement corresponding method flows or corresponding functions; the processor provided by the embodiment of the application is used for acquiring the application characteristics, processing the application characteristics, determining the application recommendation value and the like. A computer storage medium (Memory) is a Memory device in a server for storing programs and data. It is understood that the computer storage medium herein may include both a built-in storage medium of the terminal and, of course, an extended storage medium supported by the terminal. The computer storage medium provides a storage space that stores an operating system of the terminal. Also, one or more instructions, which may be one or more computer programs (including program code), are stored in the memory space and are adapted to be loaded and executed by the processor. The computer storage medium may be a high-speed RAM memory, or may be a non-volatile memory (non-volatile memory), such as at least one disk memory; and optionally at least one computer storage medium located remotely from the processor.

In the embodiment of the present application, the processor loads and executes one or more instructions stored in the computer storage medium to implement corresponding steps in the method flows provided by the embodiments in fig. 2 to fig. 7. In a specific implementation, one or more instructions in a computer storage medium are loaded by a processor and perform the following steps:

In some possible embodiments, the processor loads one or more instructions in the computer storage medium to perform the following steps:

In some possible embodiments, the first attribute feature is a first user attribute feature, and the second attribute feature is a second application attribute feature; the processor loads one or more instructions in a computer storage medium to execute an implementation manner provided by the step of determining the application recommendation network model according to the first attribute characteristics of the application to be recommended, and specifically executes the following steps:

In some possible embodiments, the sample application attribute data of each sample application includes activity level indication information; the processor loads one or more instructions in the computer storage medium to execute an implementation manner provided by the step of constructing at least one sample application feature pair according to the sample data of the at least two sample applications, and specifically executes the following steps:

constructing a positive sample characteristic i according to the first user attribute data and the sample application attribute data of the positive sample application₁And according to the first user attribute data and the negative sampleSample application attribute data of an application to construct negative sample features i₀Obtaining the sample application characteristic pair i corresponding to the sample application pair i₁₀；

In some possible embodiments, the processor loads one or more instructions in the computer storage medium to perform the implementation provided by the step of building the application recommendation network model according to the at least one sample application characteristic, and specifically performs the following steps:

In some possible embodiments, the implementation manner provided by the step of the processor loading one or more instructions in the computer storage medium to perform learning of the second attribute feature of the application to be recommended through the application recommendation network model and determining the application recommendation value corresponding to the application to be recommended specifically performs the following steps:

In some possible embodiments, any of the above sample applications corresponds to a first preset recommendation value for a positive sample in i, and a second preset recommendation value for a negative sample applicationA value; applying feature pairs i to any sample₁₀Positive sample characteristic i₁And negative sample characteristics i₀After the application recommendation network model is input, the processor loads one or more instructions in a computer storage medium to execute the following steps:

In some possible embodiments, the first attribute feature is a first application attribute feature, and the second attribute feature is a second user attribute feature; the processor loads one or more instructions in a computer storage medium to execute an implementation manner provided by the step of determining the application recommendation network model according to the first attribute characteristics of the application to be recommended, and specifically executes the following steps:

In some possible embodiments, the sample user attribute data of each sample user includes activity level indication information; the processor loads one or more instructions in the computer storage medium to execute an implementation manner provided by the step of constructing at least one sample user characteristic pair according to the user data of the at least two sample users, and specifically executes the following steps:

pairing the at least two sample users to obtain at least one sample user pair;

In some possible embodiments, the processor loads one or more instructions in the computer storage medium to perform the implementation provided by the step of building the application recommendation network model according to the at least one sample user characteristic, and specifically performs the following steps:

In some possible embodiments, the implementation manner provided by the step of the processor loading one or more instructions in the computer storage medium to perform learning of the second attribute feature of the application to be recommended through the application recommendation network model and determining the user recommendation value corresponding to the application to be recommended specifically performs the following steps:

In some possible embodiments, any sample user corresponds to a first preset recommendation value for a positive sample user in the i, and a negative sample user corresponds to a second preset recommendation value;

in any sample user feature pair i₁₀Positive sample characteristic i₁And negative sample characteristics i₀Inputting the application recommendation networkAfter the model is built, the processor loads one or more instructions in the computer storage medium to execute the following steps:

In some possible embodiments, the user attribute characteristics of the first user and/or the second user are determined by any user attribute data of user age, user gender, user study, user region, and user application account;

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by a computer program, which can be stored in a computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. The storage medium may be a magnetic disk, an optical disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), or the like.

Claims

1. A method for application recommendation, the method comprising:

acquiring application characteristics of an application to be recommended, wherein the application characteristics of the application to be recommended comprise first attribute characteristics and second attribute characteristics, the first attribute characteristics are user attribute characteristics of a first user, the second attribute characteristics are application attribute characteristics of a second application, or the first attribute characteristics are application attribute characteristics of the first application, the second attribute characteristics are user attribute characteristics of the second user, and the application attribute characteristics of the first application or the application attribute characteristics of the second application are determined by at least one of application attribute data of an application identifier, an application type, activity degree indication information and user operation habits;

determining an application recommendation network model according to the first attribute feature of the application to be recommended, learning a second attribute feature of the application to be recommended through the application recommendation network model, and determining an application recommendation value corresponding to the application to be recommended or a user recommendation value corresponding to the application to be recommended, wherein the application recommendation network model is obtained by training sample application features associated with the first user, or the application recommendation network model is obtained by training sample user features associated with the first application;

determining an application priority for recommending the application to be recommended to the first user according to the application recommendation value of the application to be recommended, or determining a user priority for recommending the first application according to the user recommendation value of the application to be recommended, wherein when the first attribute feature is a user attribute feature of a first user and the second attribute feature is an application attribute feature of a second application, the application recommendation value of the application to be recommended is an application recommendation value for recommending the second application to the first user; and when the first attribute feature is an application attribute feature of a first application and the second attribute feature is a user attribute feature of a second user, the user recommendation value of the application to be recommended is a user recommendation value for recommending the first application to the second user.

2. The method of claim 1, further comprising:

And when the user priority of the application to be recommended is greater than or equal to a preset user priority threshold, determining to recommend the first application to the second user.

3. The method according to claim 1 or 2, wherein the first attribute feature is a user attribute feature of a first user and the second attribute feature is an application attribute feature of a second application;

the determining of the application recommendation network model according to the first attribute feature of the application to be recommended includes:

matching the user attribute characteristics of the first user with the user attribute characteristics associated with each application recommendation network model in an application recommendation network model set, and determining the application recommendation network model associated with the first user corresponding to the user attribute characteristics of the first user from the application recommendation network model set;

the application recommendation network model set further includes other application recommendation network models associated with the user attribute features of other users except the first user, and the other application recommendation network models are obtained by training sample application features associated with the other users.

4. The method according to claim 3, wherein before the obtaining of the application features of the application to be recommended, the method further comprises:

obtaining sample data of at least two sample applications for application recommendation training, wherein the sample data of any sample application comprises user attribute data and sample application attribute data of the first user;

constructing at least one sample application characteristic pair according to the sample data applied by the at least two samples, wherein one sample application characteristic pair comprises a positive sample characteristic and a negative sample characteristic, the positive sample characteristic comprises the user attribute characteristic of the first user and the positive sample application attribute characteristic, and the negative sample characteristic comprises the user attribute characteristic of the first user and the negative sample application attribute characteristic;

5. The method according to claim 4, wherein the learning of the second attribute feature of the application to be recommended through the application recommendation network model, and the determining of the application recommendation value corresponding to the application to be recommended comprises:

inputting the application attribute characteristics of the second application of the application to be recommended and the user attribute characteristics of the first user into the application recommendation network model, learning the application attribute characteristics of the second application through the application recommendation network model, and outputting an application recommendation value for recommending the application to be recommended to the first user corresponding to the user attribute characteristics of the first user.

6. The method according to claim 1 or 2, wherein the first attribute feature is an application attribute feature of a first application, and the second attribute feature is a user attribute feature of a second user;

matching the application attribute characteristics of the first application with the application attribute characteristics associated with each application recommendation network model in an application recommendation network model set, and determining the application recommendation network model associated with the first application corresponding to the application attribute characteristics of the first application from the application recommendation network model set;

7. The method according to claim 6, wherein before the obtaining of the application features of the application to be recommended, the method further comprises:

acquiring user data of at least two sample users for application recommendation training, wherein the user data of any sample user comprises application attribute data of the first application and sample user attribute data;

and constructing at least one sample user characteristic pair according to the user data of the at least two sample users, and constructing an application recommendation network model according to the at least one sample user characteristic pair, wherein one sample user characteristic pair comprises a positive sample characteristic and a negative sample characteristic, the positive sample characteristic comprises the application attribute characteristic of the first application and the positive sample user attribute characteristic, and the negative sample characteristic comprises the application attribute characteristic of the first application and the negative sample user attribute characteristic.

8. The method according to claim 7, wherein the learning of the second attribute feature of the application to be recommended through the application recommendation network model, and the determining of the user recommendation value corresponding to the application to be recommended includes:

inputting the user attribute characteristics of the second user of the application to be recommended and the application attribute characteristics of the first application into the application recommendation network model, learning the user attribute characteristics of the second user through the application recommendation network model, and outputting a user recommendation value for recommending the first application to the second user corresponding to the user attribute characteristics of the second user.

9. The method according to claim 1, wherein the user attribute characteristics of the first user and/or the second user are determined by any user attribute data of user age, user gender, user study, user region and user application account.

10. A terminal, characterized in that the terminal comprises:

the system comprises a feature obtaining unit, a feature obtaining unit and a feature obtaining unit, wherein the application features of applications to be recommended comprise a first attribute feature and a second attribute feature, the first attribute feature is a user attribute feature of a first user, the second attribute feature is an application attribute feature of a second application, or the first attribute feature is an application attribute feature of the first application, the second attribute feature is a user attribute feature of the second user, and the first application attribute feature or the second application attribute feature is determined by at least one application attribute data of an application identifier, an application type, activity degree indication information and a user operation habit;

the feature processing unit is configured to determine an application recommendation network model according to the first attribute feature of the application to be recommended acquired by the feature acquisition unit, learn a second attribute feature of the application to be recommended through the application recommendation network model, and determine an application recommendation value corresponding to the application to be recommended or a user recommendation value corresponding to the application to be recommended, where the application recommendation network model is obtained by training sample application features associated with the first user, or the application recommendation network model is obtained by training sample user features associated with the first application;

a recommendation predicting unit, configured to determine, according to the application recommendation value determined by the feature processing unit, an application priority for recommending the application to be recommended to the first user, or determine, according to the user recommendation value determined by the feature processing unit, a user priority for recommending the first application, where, when the first attribute feature is a user attribute feature of a first user and the second attribute feature is an application attribute feature of a second application, the application recommendation value of the application to be recommended is an application recommendation value for recommending the second application to the first user; and when the first attribute feature is an application attribute feature of a first application and the second attribute feature is a user attribute feature of a second user, the user recommendation value of the application to be recommended is a user recommendation value for recommending the first application to the second user.

11. The terminal of claim 10, wherein the recommendation prediction unit is further configured to:

12. The terminal according to claim 10 or 11, wherein the first attribute feature is a user attribute feature of a first user, and the second attribute feature is an application attribute feature of a second application;

the feature processing unit is configured to:

matching the user attribute characteristics of the first user acquired by the characteristic acquisition unit with the user attribute characteristics associated with each application recommendation network model in an application recommendation network model set, and determining the application recommendation network model associated with the first user corresponding to the user attribute characteristics of the first user from the application recommendation network model set;

13. The terminal according to claim 10 or 11, wherein the first attribute feature is an application attribute feature of a first application, and the second attribute feature is a user attribute feature of a second user;

the feature processing unit is configured to:

matching the application attribute characteristics of the first application acquired by the characteristic acquisition unit with the application attribute characteristics associated with each application recommendation network model in an application recommendation network model set, and determining the application recommendation network model associated with the first application corresponding to the application attribute characteristics of the first application from the application recommendation network model set;

14. A computer storage medium storing instructions that, when executed on a terminal, cause the terminal to perform the method of any one of claims 1 to 9.

15. A terminal, comprising: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to perform the method according to any of claims 1-9.