CN111046298B

CN111046298B - Method and device for pushing application program, computer equipment and storage medium

Info

Publication number: CN111046298B
Application number: CN202010173116.0A
Authority: CN
Inventors: 张峻旗; 白冰; 林�也; 白琨
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2020-03-13
Filing date: 2020-03-13
Publication date: 2020-07-03
Anticipated expiration: 2040-03-13
Also published as: CN111046298A

Abstract

The application relates to a method and a device for pushing an application program based on machine learning, a computer device and a storage medium. The method comprises the following steps: acquiring an application program to be pushed, determining a pushing range according to the application category of the application program to be pushed, and acquiring an application mirror image set and an application operation record of the pushing range; inputting the application mirror image set and the application operation record into a prediction model, extracting first period demand characteristics through a first overlay network of the prediction model, extracting second period demand characteristics through a second overlay network of the prediction model, and fusing the first period demand characteristics and the second period demand characteristics through a characteristic connection layer to obtain target demand characteristics; determining interest degree scores of the application programs to be pushed of each user identification according to the target demand characteristics; and screening a target user identifier according to the interestingness score, and pushing the application program to be pushed according to the target user identifier. By adopting the method, the pushing accuracy of the application program can be effectively improved.

Description

Method and device for pushing application program, computer equipment and storage medium

Technical Field

The present application relates to the field of computer technologies, and in particular, to a method and an apparatus for pushing an application, a computer device, and a storage medium.

Background

With the rapid development of internet technology, various application software layers are in endless, such as various types of application software, such as communication software, multimedia software, office software, and the like, and a user can obtain interested application software according to needs. With the rapid development of Artificial Intelligence (AI) technology, application software can be intelligently pushed to users based on technologies such as cloud computing, distributed storage, big data processing, and the like. In the market of massive application software, some application software pushing modes appear in order to help users to quickly find interesting or needed application software.

The traditional methods are usually push methods based on logistic regression, recurrent neural network sequences or attention mechanisms, and these methods usually only carry out push according to the short-term behavior records of the user, and the short-term behavior of the user has much noise, which results in low accuracy of push. Therefore, how to effectively improve the accuracy of pushing the application software becomes a technical problem to be solved at present.

Disclosure of Invention

In view of the foregoing, it is desirable to provide a method, an apparatus, a computer device, and a storage medium for pushing an application, which can effectively improve the accuracy of pushing the application.

A method of pushing an application, the method comprising:

acquiring an application program to be pushed, wherein the application program to be pushed comprises an application category;

determining a pushing range according to the application category, and acquiring an application mirror image set and an application operation record of each user identifier corresponding to the pushing range;

inputting the application mirror image set and the application operation record into a trained prediction model, extracting a first period requirement characteristic of each user identifier through a first overlay network of the prediction model, and extracting a second period requirement characteristic of each user identifier through a second overlay network of the prediction model;

fusing the first period demand characteristics and the second period demand characteristics through a characteristic connection layer of the prediction model to obtain target demand characteristics; determining the interestingness score of each user identification corresponding to the application program to be pushed according to the target demand characteristics;

and screening a target user identifier meeting a pushing condition according to the interestingness score, and pushing the application program to be pushed to a user terminal corresponding to the target user identifier.

An apparatus for pushing an application, the apparatus comprising:

the application acquisition module is used for acquiring an application program to be pushed, and the application program to be pushed comprises an application category;

the data acquisition module is used for determining a pushing range according to the application category and acquiring an application mirror image set and an application operation record of each user identifier corresponding to the pushing range;

the interest prediction module is used for inputting the application mirror image set and the application operation record into a trained prediction model, extracting first period demand characteristics of each user identifier through a first overlay network of the prediction model, and extracting second period demand characteristics of each user identifier through a second overlay network of the prediction model; fusing the first period demand characteristics and the second period demand characteristics through a characteristic connection layer of the prediction model to obtain target demand characteristics; determining the interestingness score of each user identification corresponding to the application program to be pushed according to the target demand characteristics;

and the application pushing module is used for screening a target user identifier meeting a pushing condition according to the interestingness score and pushing the application program to be pushed to the user terminal corresponding to the target user identifier.

A computer device comprising a memory and a processor, the memory storing a computer program, the processor implementing the following steps when executing the computer program:

A computer-readable storage medium, on which a computer program is stored which, when executed by a processor, carries out the steps of:

According to the method, the device, the computer equipment and the storage medium for pushing the application program, after the server obtains the application program to be pushed, the pushing range is determined according to the application type of the application program to be pushed, and the application image set and the application operation record of each user identifier corresponding to the pushing range are obtained. The first period demand characteristics of each user identification are extracted through the first overlay network in the prediction model, the second period demand characteristics of each user identification are extracted through the second overlay network of the prediction model, and the demand characteristics of each user identification in different periods can be accurately and effectively extracted according to the application mirror image set and the application operation records. The first period demand characteristics and the second period demand characteristics are fused through a characteristic connecting layer in the prediction model, so that target demand characteristics reflecting user interests can be accurately extracted. And the server determines the interestingness score of each user identifier corresponding to the application program to be pushed by using the obtained target demand characteristics, further screens out the target user identifier meeting the pushing conditions according to the interestingness score, and pushes the application program to be pushed to the user terminal corresponding to the target user identifier. The demand characteristics of the first period and the second period of the user are extracted through the prediction model and are fused, so that the interest demand of the user can be effectively extracted comprehensively and accurately, the application program is pushed according to the interest demand of the user, and the accuracy of pushing the application program is effectively improved.

Drawings

FIG. 1 is a diagram of an application environment for a method of pushing an application in one embodiment;

FIG. 2 is a flowchart illustrating a method for pushing an application in one embodiment;

FIG. 3 is a flowchart illustrating a method for pushing an application in another embodiment;

FIG. 4 is a diagram illustrating the structure of a predictive model in one embodiment;

FIG. 5 is a flowchart illustrating a method for pushing an application in another embodiment;

FIG. 6 is a flowchart illustrating a method for pushing an application in accordance with yet another embodiment;

FIG. 7 is a flowchart illustrating a method for pushing an application in accordance with another embodiment;

FIG. 8 is a diagram that illustrates an interface for an application push list, according to an embodiment;

FIG. 9 is a schematic flow chart diagram illustrating the predictive model training step in one embodiment;

FIG. 10 is an overall framework diagram of the application push process in one embodiment;

FIG. 11 is a block diagram of an apparatus for pushing an application in one embodiment;

FIG. 12 is a block diagram showing an arrangement of pushing an application in another embodiment;

FIG. 13 is a block diagram showing an arrangement of pushing an application in another embodiment;

FIG. 14 is a diagram illustrating an internal structure of a computer device according to an embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.

The method for pushing the application program can be applied to the application environment shown in fig. 1. Wherein the server 110 communicates with the user terminal 120 through a network. After obtaining the application program to be pushed, the server 110 determines a pushing range according to the application category of the application program to be pushed, obtains an application image set and an application operation record of each user identifier corresponding to the pushing range, and inputs the application image set and the application operation record to the trained prediction model. The first period demand characteristics of each user identification are extracted through a first stack network of the prediction model, the second period demand characteristics of each user identification are extracted through a second stack network of the prediction model, the server 110 fuses the first period demand characteristics and the second period demand characteristics through a characteristic connection layer of the prediction model to obtain target demand characteristics, interestingness scores of application programs to be pushed corresponding to each user identification are determined according to the target demand characteristics, then target user identifications meeting pushing conditions are screened according to the interestingness scores, and the application programs to be pushed are pushed to a plurality of user terminals 120 corresponding to the target user identifications. The server 110 may be implemented by an independent server or a server cluster composed of a plurality of servers, and the terminal 120 may be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.

In one embodiment, as shown in fig. 2, a method for pushing an application program is provided, which is described by taking the method as an example for being applied to the server 110 in fig. 1, and includes the following steps:

s202, acquiring an application program to be pushed, wherein the application program to be pushed comprises an application category.

The application program is one of the main categories of computer software, and refers to application software written for a certain application purpose of a user. The types of the application programs include various types, and for example, the application programs can include various types such as instant messaging software, enterprise messaging software, music software, image processing software, office software, various types of game software and the like. Each application program includes a corresponding application category and application attribute information. The application attribute information may include information such as version identification and application range of the application program.

After the application program is developed, a developer corresponding to the application program can be released to the application program pushing system to push the application program to the user. The application pushing system can comprise a plurality of applications to be pushed. For example, an application developer may initiate a placement request or an advertisement request, etc. to an application push system. The application pushing system can be an application system based on artificial intelligence and machine learning, and the application pushing system is used for conducting prediction analysis on the application program and the target user, so that the application program is pushed to users meeting the conditions.

The artificial intelligence is a theory, a method, a technology and an application system which simulate, extend and expand human intelligence by using a digital computer or a machine controlled by the digital computer, sense the environment, acquire knowledge and obtain the best result by using the knowledge. For example, technologies such as cloud computing, distributed storage, big data processing, machine learning and the like can be adopted to enable the machine to have functions of perception, reasoning and decision making. Therefore, the application pushing system has the functions of reasoning and decision making, so that the application program meeting the requirements of the user can be intelligently pushed to the user.

S204, determining a pushing range according to the application type, and acquiring an application mirror image set and an application operation record of each user identifier corresponding to the pushing range.

Wherein the push scope represents the application scope of the application program to be pushed, for example, some learning-class application programs are applicable to users in a specific age group. The mirror image is a file storage form, a duplicate of data on one disk on another disk is the mirror image, and a mirror image file can be generated by a specific series of files according to a certain format. The application installation image refers to an image file corresponding to an application program installed in the user terminal and is used for representing the application program currently installed in the user terminal. The application image set refers to a set of all application programs currently installed in the user terminal.

The application operation record refers to a record in which a user operates with respect to an application program in the user terminal, for example, a record of installing the application program and uninstalling the application program. The application operation record comprises an application identification, an operation type and an operation time, wherein the operation type comprises an installation operation and an uninstallation operation.

In one embodiment, the application operation record comprises an application identification, an operation type and an operation time; the operation type includes at least one of an application install operation and an application uninstall operation.

The application operation record comprises complete record information of each operation of the application program by a user, and the application operation record comprises an application identifier, an operation type and operation time; the operation type includes at least one of an application installation operation and an application uninstallation operation, that is, the operation of the user on each application program may include only the installation operation or may include the installation operation and the uninstallation operation.

And after the server acquires the application program to be pushed, determining a target pushing range according to the application category of the application program to be pushed. Specifically, the server may determine a pushing range according to the application category and the application attribute information of the application program to be pushed, determine users meeting the conditions according to the pushing range, and obtain an application image set and an application operation record corresponding to the user identifiers.

And S206, inputting the application mirror image set and the application operation record into the trained prediction model, extracting the first period requirement characteristics of each user identifier through the first overlay network of the prediction model, and extracting the second period requirement characteristics of each user identifier through the second overlay network of the prediction model.

The period is a period with certain characteristics and length in the process of representing the development of the object, and the requirement characteristic can be embodied as the installation requirement of the user on the application program. The first period is longer than the second period, and the first period requirement characteristic may be a first period requirement characteristic, and the first period requirement characteristic represents a requirement of the user for the application program for a longer period, for example, the longer period may be defined by a preset time length of one month, three months, six months or more. The second period requirement characteristic may be a second period requirement characteristic, and the second period requirement characteristic represents a requirement of the user for the application program in a shorter period, for example, the shorter period may be defined according to a preset time length of one month, two weeks, less than one week, and the like. E.g. some long-existing applications in the user terminal, indicate that the user has a long-term need for these applications. The second period requirement characteristic represents the short-term requirement of the user for the application programs, such as the application programs which are installed in the user terminal but are unloaded in the short term, namely represents the specific short-term requirement of the user for the user programs.

Machine Learning (ML) is a multi-domain cross discipline, and relates to a plurality of disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer.

The prediction model can be a neural network model based on machine learning, the prediction model is a neural network model with prediction capability obtained by a server through pre-training, and the prediction model is obtained by training an application image set, an application operation record and training label data of a large number of users. The predictive model may include a multi-layer neural network structure for extracting long and short term needs or interests of the user from the set of application images and application operational records to push the application to the user.

In particular, the predictive model includes a plurality of overlay networks, which may include a multi-layer neural network structure. The prediction model may include a first overlay network and a second overlay network, and the first overlay network and the second overlay network are different stacked neural network layers. The first overlay network may be an overlay automatic coding machine, which refers to a kind of artificial neural network structure, and encodes the network input by reconstructing the network input, and completes information extraction at the same time. For example, an SdA Stacked automatic coding machine (Stacked automatic coding machine) may be used. The laminated automatic coding machine can be an automatic coding machine comprising a plurality of intermediate hidden layers, and each layer is provided with corresponding weight parameters.

The second laminated network can be a laminated encoder which comprises a plurality of automatic encoding layers with hidden layers in the middle, and the laminated encoder is an artificial neural network structure based on a multi-head self-attention mechanism and a feedforward neural network, supports the laminated structure and has stronger information extraction capability. For example, a transform encoder, each layer is provided with a corresponding weight parameter.

The server acquires application mirror image sets and application operation records of a plurality of user identifications according to the target push range, inputs the application mirror image sets and the application operation records of the user identifications into a prediction model, and performs feature extraction on the application mirror image sets and the application operation records through the prediction model to obtain first period demand features and second period demand features of the user identifications. Specifically, the server extracts a first period requirement characteristic of each user identifier through a first overlay network of the prediction model, and extracts a second period requirement characteristic of each user identifier through a second overlay network of the prediction model. The characteristic extraction is carried out on the application mirror image set and the application operation records through the multi-layer neural network structure in the optimal model, and the requirement characteristics of each user in different periods can be accurately extracted.

S208, fusing the first period demand characteristics and the second period demand characteristics through a characteristic connection layer of the prediction model to obtain target demand characteristics; and determining the interestingness score of each user identifier corresponding to the application program to be pushed according to the target demand characteristics.

The prediction model further comprises a feature connection layer and a feedforward neural network layer, and the feature connection layer is used for performing feature fusion on the first period demand feature and the second period demand feature which are extracted by the first overlay network and the second overlay network respectively. And the feed-forward neural network layer is used for predicting the interestingness score of the application program to be pushed corresponding to the user identification according to the target demand characteristics output by the characteristic connection layer.

The interestingness score may reflect the user's interestingness in the application, and the interestingness score may be a predicted probability value of the application that the user will install or continue to retain in the future.

After the server extracts the first-period demand characteristics and the second-period demand characteristics of each user identifier through the prediction model, the obtained first-period demand characteristics and the obtained second-period demand characteristics are input into the characteristic connection layer in the prediction model to be fused, and therefore the target demand characteristics fused with the user in different periods can be obtained. For example, the first period demand characteristics may be long-term demand characteristics, and the second period demand characteristics may be short-term demand characteristics, so that a user interest demand combining the long-term demand and the short-term demand can be obtained, and thus, target demand characteristics which reflect user interest more truly and comprehensively can be accurately and effectively extracted.

And after the server performs feature fusion on the first-period demand features and the second-period demand features of the user identifiers through the prediction model, the server further calculates the interestingness score of each user identifier for the application program to be pushed according to the fused target demand features. Specifically, the server can predict the application categories in which the user is interested based on the fused user interest features through a feedforward neural network layer in the prediction model, and obtains the interestingness scores of the application programs to be pushed corresponding to the user identifiers based on the prediction probability values of the application categories, so that comprehensive and accurate user interest requirements can be effectively obtained according to the requirement features in different periods.

S210, screening the target user identification meeting the pushing condition according to the interestingness score, and pushing the application program to be pushed to the user terminal corresponding to the target user identification.

And after calculating the interestingness score of each user identifier corresponding to the application program to be pushed through the prediction model, the server extracts the target user identifier of which the interestingness score reaches the condition threshold, and generates a user pushing list by using the user identifier of which the interestingness score meets the condition threshold. And the server pushes the application program to be pushed to the user terminal corresponding to each user identifier in the user pushing list.

In one embodiment, after the server screens out a plurality of target user identifiers according to the interestingness score, a user push list can be generated by using the plurality of target user identifiers. The user push list is a user list which is required to be pushed to a plurality of users aiming at the application program to be pushed, and the user push list comprises a plurality of user identifications meeting the push conditions. And the server pushes the application program to be pushed to the user terminal corresponding to each user identifier in the user pushing list.

In one embodiment, the server may obtain a plurality of applications to be pushed simultaneously, and predict corresponding user push lists respectively by using the prediction model, where the user push lists of the plurality of applications to be pushed may include overlapping user identifiers. The user terminal corresponding to each user identifier can also receive the application programs pushed by a plurality of servers at the same time. Specifically, after receiving a plurality of pushed application programs, the user terminal sorts the interestingness scores of the application programs according to the user identification, generates a sorted application push list, and displays the application push list on a display interface of the user terminal.

In the method for pushing the application program, after the server acquires the application program to be pushed, the server determines a pushing range according to the application category of the application program to be pushed, and acquires an application mirror image set and an application operation record of each user identifier corresponding to the pushing range. The first period demand characteristics of each user identification are extracted through the first overlay network in the prediction model, the second period demand characteristics of each user identification are extracted through the second overlay network of the prediction model, and the demand characteristics of each user identification in different periods can be accurately and effectively extracted according to the application mirror image set and the application operation records. The first period demand characteristics and the second period demand characteristics are fused through a characteristic connecting layer in the prediction model, so that target demand characteristics reflecting user interests can be accurately extracted. The server determines the interestingness score of each user identification corresponding to the application program to be pushed according to the obtained target demand characteristics, screens the target user identification meeting the pushing conditions according to the interestingness score, and pushes the application program to be pushed to the user terminal corresponding to the target user identification. The first period demand characteristics and the second period demand characteristics of the user are extracted through the prediction model and are fused, so that the interest demands of the user can be acquired comprehensively and accurately effectively, the application program is pushed according to the interest demands of the user, and the accuracy of pushing the application program is improved effectively.

In an embodiment, as shown in fig. 3, a method for pushing an application program is provided, which specifically includes:

s302, acquiring an application program to be pushed, wherein the application program to be pushed comprises an application category.

S304, determining a target pushing range according to the application type, and acquiring an application mirror image set and an application operation record of each user identifier corresponding to the target pushing range.

S306, inputting the application mirror image set and the application operation record into the trained prediction model, and extracting the first time period demand characteristics of each user identifier according to the application mirror image set by using the first overlay network of the prediction model.

After the server acquires the application program to be pushed, a target pushing range is determined according to the application category of the application program to be pushed, and an application mirror image set and an application operation record of each user identifier corresponding to the target pushing range are acquired. And inputting the application mirror image set and the application operation record of each user identifier into a trained prediction model, and performing feature extraction on the application mirror image set and the application operation record through the prediction model.

The prediction model comprises a multi-layer neural network structure, wherein the multi-layer neural network structure comprises a first overlay network and a second overlay network. The first overlay network may be an overlay transcoder layer including a plurality of intermediate hidden layers for extracting first time-period requirement characteristics of respective subscriber identities. The second overlay network may be an overlay encoder comprising a plurality of intermediate hidden layers for extracting second epoch requirement characteristics for respective user identities.

The application image set comprises a plurality of application installation images currently installed in the user terminal, and the application image set can reflect long-term interest demands of users on the application programs currently in use. The application operation records comprise record information such as application installation operation, application uninstallation operation and the like, and can reflect short-term interest and demands of users on the application programs with operation behaviors.

Specifically, after the application image set and the application operation record are input to the prediction model, the server inputs the application installation images in the application image set to a first overlay network of the prediction model, and performs feature extraction on the application installation images in the image set through the first overlay network, so that first time-period demand features reflecting long-term interest demands of users can be extracted. Taking the first overlay network as an overlay automatic coding layer as an example, reconstructing and coding the application installation mirror image through the overlay automatic coding layer, so as to extract the first time period requirement characteristics of the user identifier.

And S308, extracting second-period requirement characteristics of each user identification according to the application operation records by utilizing the second overlay network of the prediction model.

And the server simultaneously inputs the application operation records of the user identifications to a second overlay network in the prediction model, and the second overlay network extracts second period requirement characteristics of each user identification according to the application operation records. For example, the second overlay network may be an overlay encoder, and the application installation feature representation and the application uninstallation representation are extracted layer by layer through the overlay encoding layer, so as to extract the corresponding second-period requirement feature according to the installation feature representation and the application uninstallation representation.

And S310, performing feature fusion on the first period requirement features and the second period requirement features through a feature connection layer of the prediction model to obtain target requirement features of each user identifier.

After the server respectively extracts the first period demand characteristics and the second period demand characteristics of each user identification through the prediction model, the first period demand characteristics and the second period demand characteristics are further input into a characteristic connection layer in the prediction model, the first period demand characteristics and the second period demand characteristics are fused through the characteristic connection layer, the target demand characteristics of each user identification are extracted through the first period demand characteristics and the second period demand characteristics, the target demand characteristics reflect the comprehensive interest requirements of the users, and therefore the interest characteristics of each user can be accurately extracted.

And S312, determining the interestingness score of each user identifier corresponding to the application program to be pushed according to the target demand characteristics.

And S314, screening the target user identification meeting the pushing condition according to the interestingness score, and pushing the application program to be pushed to the user terminal corresponding to the target user identification.

The server further calculates the interestingness score of each user identifier for the application program to be pushed according to the fused target demand characteristics, can predict the application category interested by the user based on the fused target demand characteristics of each user through a feedforward neural network in the prediction model, and obtains the interestingness score of the application program corresponding to the user identifier based on the prediction probability value of the application category. And the server extracts the user identification of which the interestingness score reaches the condition threshold, and generates a user push list by using the user identification of which the interestingness score meets the condition threshold.

In the embodiment, after the first period demand characteristics and the second period demand characteristics of each user identifier are respectively extracted, the first period demand characteristics and the second period demand characteristics are fused, and the target demand characteristics of each user identifier are extracted by using the first period demand characteristics and the second period demand characteristics, so that the interest demands of users can be accurately and effectively extracted, the application program is pushed according to the interest demands of the users, and the accuracy of pushing the application program is effectively improved.

In one embodiment, extracting the first time period requirement characteristics of each user identification according to the application image set by using the first overlay network of the prediction model comprises: inputting the application mirror image set into a first overlay network, and performing feature extraction on an application installation mirror image in the application mirror image set to obtain corresponding application mirror image features; coding the application mirror image characteristics by using the first overlay network and the corresponding weight parameters to obtain coded application mirror image characteristics; and extracting the first time period demand characteristics of each user identifier according to the coded application mirror image characteristics by using a preset function.

After the application mirror image set and the application operation record are input to the prediction model by the server, the server firstly inputs the application mirror image set to a first overlay network in the prediction model in the process of performing feature extraction on the application mirror image set by using the prediction model. The first overlay network firstly performs feature extraction on the application mirror image set to obtain corresponding application mirror image features. The server further encodes each application installation mirror image in the application mirror image set layer by utilizing each layer of network of the first laminated network and the weight parameters of each layer of network, so as to obtain the encoded application mirror image characteristics. The prediction model extracts the first time-period demand characteristics of each user identification according to the coded application image characteristics by using a preset function, so that the interest demands corresponding to the first time-period attributes of the users can be accurately extracted according to the application image set.

In particular, the first time period demand characteristic may be a long-term demand characteristic of the user. The set of application images of all users of the target push scope may be represented by a matrix

Indicating that where U is the number of users and M is the number of applications. The user's application image set is available

Is shown asx _umAnd 1, it indicates that the application m is currently installed on the user terminal of the user u. The application operation record may be as follows:

wherein the content of the first and second substances,

an application representing the ith new load of user u,

which indicates the time of the corresponding operation,

and

for indicating the corresponding uninstall sequence of the application.

And after the server inputs the application image sets and the application operation records of the plurality of user identifications into the prediction model, the application image sets are input into a laminated automatic coding machine in the prediction model to be reconstructed and coded, so that the first time requirement characteristics of the user are extracted.

The prediction model may adopt a stacked automatic coding machine with three intermediate hidden layers, which has 4 layers of weighting parameters in total, and the formula of the stacked automatic coding machine may be specifically expressed as:

wherein, based on the useruUser installed mirror imagex _u，W ^(p)Is as followspThe weight of a layer is determined by the weight of the layer,W ⁽¹⁾may be the weight of the first layer;b ^(p)is as followspThe biasing of the layers is such that,x ^(p)andx ^(p-1)are respectively the firstpThe output and the input of the layer are,f ^(p)is shown aspThe activation function of the layer.x ⁽⁰⁾I.e. mounting the mirror image for the input of the automatic coding machinex _u，x ⁽⁴⁾I.e. installing the image for the reconstructed user,x ⁽¹⁾、x ⁽²⁾、x ⁽³⁾the user installation images are coded and can be used for representing long-term requirement characteristics of the user. Specifically, the server may reconstruct the application installation mirror features using a cross entropy loss function to improve the accuracy of feature reconstruction.The loss function is specifically as follows:

and after the user installation mirror image of each layer is reconstructed and coded through a loss function or an activation function, the long-term requirement characteristics of the user can be accurately extracted.

In one embodiment, extracting second age requirement characteristics for each user identification from the application operation record using the second overlay network of the predictive model comprises: inputting the application operation record into a second overlay network, and performing feature extraction on the application operation record to obtain application embedded representation, operation time representation and operation type representation; coding the application embedding representation, the application operation representation and the operation time representation by utilizing the second overlay network and corresponding weight parameters to obtain an application operation recording matrix; and extracting the second-period requirement characteristics of each user identifier according to the application operation recording matrix.

The application operation record comprises an application identifier, an operation type and operation time corresponding to each operation of the application program by a user; the operation type includes at least one of an application installation operation and an application uninstallation operation, that is, the operation of the user on each application program may include only the installation operation or may include the installation operation and the uninstallation operation.

The second overlay network in the prediction model may specifically be an overlay encoder. And after the application operation record is input into the prediction model, the server extracts an application operation record vector corresponding to the application operation record, and the application operation record vector is used as the input of the laminated encoder. Firstly, the embedded layer of the laminated encoder respectively extracts the characteristics of the application operation record to obtain an application embedded representation, an operation time representation and an operation type representation. And further encoding the application embedding representation, the application operation representation and the operation time representation by using the second overlay network and corresponding weight parameters to obtain an application operation recording matrix. And the second period requirement characteristics of the user can be further accurately extracted by using the application operation record matrix.

In one embodiment, extracting the second period requirement characteristics of each user identifier according to the application operation record matrix comprises: acquiring application mirror image characteristics corresponding to the application mirror image set; inputting the application mirror image characteristics into an application operation record matrix for characteristic fusion to obtain a fused application matrix; and extracting the second-period demand characteristics of each user identifier according to the application matrix.

Wherein the second period demand characteristic may be a short-term demand characteristic of the user. Due to the fact that noise may exist in the application operation record of the user, for example, there is a possibility that the user performs misoperation or careless operation in the application operation record, such noise may affect the accuracy of short-term user demand extraction, and thus the accuracy of recommendation is affected.

In the process that the server extracts the short-term demand of the user through the second overlay network of the prediction model by using the application operation record, the server can also input the application image features corresponding to the application image set into the application operation record matrix for feature fusion to obtain a fused application matrix, and then extract the second-period demand features of each user identifier according to the application matrix, so that the short-term demand features of the user with high accuracy can be effectively extracted.

Specifically, after extracting the application embedded representation of the application identifier in the application operation record through the overlay encoder, adding the operation type representation for encoding on the basis of the application embedded representation, and obtaining the application installation characteristic representation and the application uninstallation characteristic representation. And further adding operation time representation to encode, and obtaining operation record representation of each application, thereby obtaining an application operation record matrix corresponding to a plurality of operation record representations based on the user. And calculating the application operation recording matrix through a preset function in the laminated encoder, so that the short-term requirement characteristics of the user can be obtained.

For example, the second overlay network may be embodied as a transform Encoder, which is a model based on an Encoder-Decoder (Encoder-Decoder) structure. The Encoder and Decoder may be stacked from a plurality of network layers, for example, including at least 3 layers, each of which may further include a corresponding sub-layer. Each layer has a corresponding weight parameter.

Fig. 4 is a schematic structural diagram of a prediction model in a specific embodiment. L as shown in the schematic diagram of the prediction model structure_recThe corresponding structural part can represent a network structural layer for processing the application image set of the user by adopting a laminated automatic coding machine; the structural parts corresponding to the Encodinglayer and the Stacking can represent the network structural layer for processing the application operation record by adopting a laminated Transformer encoder; the Concat part can represent a feature connection layer and is used for fusing the extracted first period requirement features and the extracted second period requirement features; l is_preThe corresponding structural part can be a feedforward neural network and is used for carrying out prediction processing according to the fused demand characteristics.

The transform encoder may include a multi-head self-attention-mechanism Layer (multihead self-attention-mechanism), a position-wise full link Layer (position-wise fed Forward), an intermediate Layer may include a multilayer residual error Normalization Layer (Layer Normalization) and an FFN (Feed Forward neural network), and each sub-Layer has a residual error connection for Layer Normalization.

Specifically, taking the first period requirement characteristic as the long-term requirement characteristic of the user and the second period requirement characteristic as the short-term requirement characteristic of the user as an example, the application operation record based on the user u includes records of application installation and application uninstallation of the user. Specifically, the short-term demand characteristics of the user can be extracted and fused with the long-term demand of the user through the following steps.

Step 1: the application installation and application uninstallation records are converted into corresponding application embedded representations by taking the application operation record vector as the input of the overlay encoder. Suppose that

The application embedding of the strip of records is represented as

。

Step 2: obtained by conversion in the previous stepAs a result, the add operation type embedded representation is encoded. Wherein the mounting features are represented asW ⁿApplication offload features are represented asW ^l。

And step 3: on the result converted out in the previous step, an operation time representation is added. Suppose that

The operation time of the application operation record is expressed as

。

Through the three steps, the application installation and uninstallation operation record at any time can be converted into an application operation record matrix as follows, namely the embedded layer representation containing the application representation, the operation time and the operation type:

and 4, step 4: acquiring the encoded application installation image, and adding an embedded layer representation of the application installation image, as follows:

wherein in this step, ax ⁽¹⁾The image is installed as an encoded application.

Step 5; the input to the stacked transform encoder is collated. All embedded layer representations obtained in the first 4 steps are sorted into a list as shown below:

step 6: and processing the application operation record matrix converted into the embedded layer representation and the coded application installation image by using an overlay Transformer encoder so as to fuse the application image characteristics in the application image set into the application operation record. The single-layer transform encoder is specifically shown below:

wherein, MultiHead represents a multi-head self-attention mechanism, LayerNorm represents layer normalization, and FFN represents a feedforward neural network;e ⁽⁰⁾indicates any one of the embedded layer representation lists obtained in step 5,e ^(r-1)represents the input of the r-th layer and corresponds to any one of the embedded layer representation lists obtained in step 5.

And 7: and fusing the application mirror image features in the application mirror image set to the application operation record matrix, thereby extracting and obtaining short-term demand feature representation fused with the long-term demand features of the user. E.g., common R layer transform encoder, will

As short-term user requirements extracted from the application operation record. The application mirror image set and the application operation records are subjected to feature fusion extraction through the laminated encoder, so that the short-term requirement features of the user with high accuracy can be effectively extracted.

In one embodiment, fusing the first term demand characteristics and the second term demand characteristics through the characteristic connection layer of the predictive model comprises: inputting the first period demand characteristics and the second period demand characteristics into a characteristic connecting layer, and splicing the first period demand characteristics and the second period demand characteristics to obtain spliced demand characteristics; and generating target demand characteristics of each user identification by using the spliced demand characteristics.

The prediction model comprises a first overlay network, a second overlay network and a feature connection layer, and is used for performing feature fusion on a first period demand feature and a second period demand feature which are extracted by the first overlay network and the second overlay network respectively.

The server inputs the application mirror image set and the application operation records of each user identifier into the prediction model, performs feature extraction on the application mirror image set through the prediction model to obtain first period demand features, and performs feature extraction on the application operation records to obtain second period demand features, and then the server further fuses the first period demand features and the second period demand features through the prediction model.

Specifically, the server performs feature extraction on the application mirror image set through a first overlay network of the prediction model to obtain a first period demand feature, and performs feature extraction on the application operation record through a second overlay network to obtain a second period demand feature; and inputting the first period demand characteristic and the second period demand characteristic into the characteristic connecting layer, and splicing the first period demand characteristic and the second period demand characteristic to obtain spliced demand characteristics. The server generates target demand characteristics of each user identification through a preset function by using the spliced demand characteristics, so that the interest requirements of the users can be effectively extracted and obtained comprehensively and accurately.

And the server further determines the interestingness score of each user identifier corresponding to the application program to be pushed by using the target demand characteristics of the user, further generates a user pushing list by using the target user identifier of which the interestingness score meets a condition threshold, and pushes the application program to be pushed to the user terminal corresponding to the user pushing list. The first period demand characteristics and the second period demand characteristics of the user are extracted through the prediction model and are fused, so that the interest requirements of the user can be accurately and effectively extracted, the application program is pushed according to the interest requirements of the user, and the accuracy of pushing the application program is effectively improved.

Specifically, after the long-term requirement characteristics and the short-term requirement characteristics of the users are extracted through the overlay automatic encoding machine and the overlay Transformer encoder, and then the process of recommending the application program by using the long-term and short-term requirement characteristics of each user may be as follows:

the server extracting the laminated automatic coding machineLong term demand characteristicsx ⁽²⁾And short-term demand features extracted by a stacked transform encoder

And inputting the feature into the feature connection layer for feature splicing, so that the long-term and short-term demand features which are combined with the long-term demand and the short-term demand of the user can be obtained. The expression for fusing the long-term demand characteristics and the short-term demand characteristics may be as follows:

wherein the content of the first and second substances,

the requirements characteristic resulting from the splicing can be represented,

a fused long and short term demand representation can be represented.

And the server further predicts the interestingness score of the user on the application program to be pushed through a preset loss function by utilizing a feedforward neural network in the prediction model. Specifically, a cross entropy loss function can be used for prediction to improve the prediction accuracy of the user on the installation behavior of the application program, where the loss function is specifically as follows:

wherein the content of the first and second substances,L _prethe probability that the user will install and retain various categories of applications for a future period of time may be expressed. The long-term and short-term demand characteristics can reflect the interest demands of the users, and the application programs are pushed according to the interest demands of the users, so that the accuracy of pushing the application programs is effectively improved.

In one embodiment, the method further comprises: calculating the interestingness score of each user identifier corresponding to the application category of the application program to be pushed according to the target demand characteristics; generating a user push list corresponding to each application category according to the interestingness score; and pushing the application programs to be pushed of each application category to the corresponding user terminals according to the user pushing list.

The server extracts first-period demand characteristics and second-period demand characteristics of each user through the prediction model, fuses the first-period demand characteristics and the second-period demand characteristics, and generates target demand characteristics of the users by using the fused target demand characteristics. And in the process that the server determines the interestingness score of the application program to be pushed corresponding to each user identifier by using the target demand characteristics of each user, the application category of the application program is used as a prediction target. Namely, the server may calculate the interestingness score of each user corresponding to the application category according to the application category of the application program to be pushed. Therefore, when calculating the interestingness score of the user on the application program to be pushed, only the interestingness score of the user on the application category of the application program to be pushed needs to be calculated.

If there are multiple to-be-pushed application programs of the same application category, the interestingness scores of the to-be-pushed application programs corresponding to the user may also be the same. And the server generates a user push list corresponding to each application category according to the interestingness score, wherein each application category can predict to obtain a corresponding user push list, and the user push lists of each application category are different. And the server further pushes the application programs to be pushed of each application category to the corresponding user terminals according to the user pushing list. The interestingness score of the user is predicted according to the application category of the application program, so that the user can be pushed quickly and accurately according to the application category, and the pushing accuracy and the pushing efficiency of the application program can be effectively improved.

As shown in fig. 5, in a specific embodiment, the pushing method of the application program includes the following steps:

s502, acquiring the application program to be pushed, wherein the application program to be pushed comprises the application category.

S504, determining a pushing range according to the application type, and acquiring an application mirror image set and an application operation record of each user identifier corresponding to the pushing range.

S506, inputting the application image set and the application operation records into the trained prediction model.

And S508, inputting the application mirror image set into a first overlay network of the prediction model, and performing feature extraction on the application installation mirror images in the application mirror image set through the first overlay network to obtain corresponding application mirror image features.

S510, the application mirror image characteristics are coded by the first overlay network and the corresponding weight parameters, and the coded application mirror image characteristics are obtained.

And S512, extracting the first time period requirement characteristics of each user identifier according to the coded application mirror image characteristics by using a preset function.

And S514, inputting the application operation record into a second overlay network of the prediction model, and performing feature extraction on the application operation record through the second overlay network to obtain an application embedded representation, an operation time representation and an operation type representation.

And S516, encoding the application embedding representation, the application operation representation and the operation time representation by using the second overlay network and corresponding weight parameters to obtain an application operation recording matrix.

And S518, extracting the second period requirement characteristics of each user identifier according to the application operation recording matrix.

And S520, inputting the first period demand characteristics and the second period demand characteristics into a characteristic connection layer of the prediction model, and splicing the first period demand characteristics and the second period demand characteristics through the characteristic connection layer to obtain target demand characteristics of each user identifier.

And S522, determining the interestingness score of each user identifier corresponding to the application program to be pushed according to the target demand characteristics.

And S524, screening the target user identification meeting the pushing condition according to the interestingness score, and pushing the application program to be pushed to the user terminal corresponding to the target user identification.

According to the pushing method of the application program, the first overlay network and the second overlay network of the prediction model are used for respectively extracting the characteristics of the application mirror image set and the application operation record, and the first period requirement characteristics and the second period requirement characteristics of each user identification can be accurately and effectively extracted. The first period requirement characteristics and the second period requirement characteristics are fused, so that the target requirement characteristics which meet the user interest can be accurately and comprehensively extracted. The server determines the interestingness score of each user identification corresponding to the application program to be pushed by using the fused demand characteristics, further generates a user pushing list by using the target user identification of which the interestingness score meets the condition threshold, and pushes the application program to be pushed to the user terminal corresponding to the user pushing list. The first period demand characteristics and the second period demand characteristics of the user are extracted through the prediction model and are fused, so that the more comprehensive and accurate interest demands of the user can be effectively obtained, the application program is pushed according to the interest demands of the user, and the accuracy of pushing the application program is effectively improved.

As shown in fig. 6, in one embodiment, there is provided a pushing method of an application program, including the following steps:

s602, receiving an application acquisition request sent by a user terminal, wherein the application acquisition request carries a user identifier.

S604, acquiring a current application mirror image set and an application operation record of the user identification, and inputting the current application mirror image set and the application operation record into the prediction model to obtain the interestingness score of each application category corresponding to the user identification.

And S606, extracting the application categories meeting the conditions according to the interestingness scores, and acquiring the application programs to be pushed corresponding to the application categories.

S608, pushing the application program to be pushed to the user terminal.

Wherein the current application image set represents a set of all applications currently installed in the user terminal. The application operation record represents the operation record of the user to the application program in a historical period of time, and comprises an installation record and an application uninstallation record. For example, the application management software may include an application push system, and the application push system includes a plurality of applications to be pushed. The application to be pushed may represent an application that has been published in the application push system but that has not been installed by the user.

The user terminal may also initiate an application acquisition request to the server, for example, when the user accesses the application management software, the application acquisition request may be initiated to the server. And after receiving an application acquisition request sent by the user terminal, the server acquires a current application mirror image set and an application operation record in the user terminal according to the user identification.

The server inputs the current application mirror image set and the application operation record of the user into the prediction model, and the first time demand characteristic of the user can be obtained by performing characteristic extraction on the current application mirror image set of the user through the first overlay network of the prediction model. And simultaneously, carrying out feature extraction on the application operation record of the user through a second overlay network to obtain a second period requirement feature of the user. And the server performs feature fusion on the first period demand feature and the second period demand feature of the user through the prediction model to obtain the target demand feature of the user.

The server further obtains application categories of a plurality of application programs to be pushed in the application pushing system, calculates interestingness scores of the user identifications corresponding to the application categories according to target demand characteristics of the users, and obtains the application categories of which the interestingness scores reach a condition threshold. Wherein, each application category may comprise a plurality of applications to be pushed. The server may obtain a preset number of applications to be pushed with the highest ranking under each application category. And the server acquires the preset number of the application programs to be pushed under the application types meeting the conditions and pushes the acquired application programs to be pushed to the user terminal.

In the embodiment, based on an application acquisition request initiated by a user, a current application mirror image set and an application operation record of the user are analyzed through a prediction model, a first period requirement characteristic and a second period requirement characteristic of the user are extracted and fused, so that a target requirement characteristic which is relatively comprehensive and accurate and meets the user interest can be accurately and effectively extracted, the application type which is interested by the user is predicted, a corresponding application program to be pushed is obtained for pushing, the application program can be accurately pushed according to the interest requirement of the user, and the pushing accuracy of the application program is effectively improved.

As shown in fig. 7, in one embodiment, there is provided a push method of an application program, including the following steps:

s702, receiving an application acquisition request sent by a user terminal, wherein the application acquisition request carries a user identifier.

S704, acquiring a current application mirror image set and an application operation record of the user identification, and inputting the current application mirror image set and the application operation record into the prediction model to obtain interestingness scores of the user identification corresponding to each application category.

S706, extracting the application categories meeting the conditions according to the interestingness scores, and acquiring the application programs to be pushed corresponding to the application categories.

S708, the heat value of each application program to be pushed corresponding to the application type is obtained.

And S710, extracting the application program to be pushed with the heat value meeting the condition threshold.

And S712, generating an application push list according to the obtained application program to be pushed and the interestingness score.

And S714, pushing the application program to be pushed in the application pushing list to the user terminal.

The server extracts the target demand characteristics of the user according to the current application mirror image set and the application operation records of the user through the prediction model, calculates the interestingness score of each application category corresponding to the user identification according to the target demand characteristics of the user, and then obtains the application category of which the interestingness score reaches the condition threshold. Each application category may include a plurality of applications to be pushed, each application to be pushed includes corresponding application information, for example, information such as a download rate and a click rate, and the download rate and the click rate of the application may reflect the popularity of the application.

Specifically, the server obtains an application program to be pushed in the application category meeting the condition, wherein the application program to be pushed comprises application information. And the server acquires the heat value of each application program to be pushed according to the application information of the application program to be pushed. The server can directly determine the corresponding heat value according to the download rate of the application program to be pushed, and can also calculate the corresponding heat value according to the download rate and the click rate of the application program to be pushed. And the server extracts the application programs to be pushed with the heat values meeting the condition threshold value under each application category, and generates an application pushing list by using the acquired application programs to be pushed.

The server can also sequence the acquired application programs to be pushed according to the interestingness scores, and in the sequencing process, the interestingness scores can be used as a first weight, and the heat values can be used as a second weight. The application programs to be pushed are sorted according to the interestingness scores, and then the application programs to be pushed are sorted according to the popularity values, so that a sorted application pushing list is obtained. Fig. 8 is a schematic interface diagram of an application push list presented to a user terminal. And the server pushes the sequenced application push list to the user terminal. Therefore, the application programs which are interested by the user and are popular can be displayed in the interface of the user terminal, so that the installation probability of the user to the pushed application programs can be effectively improved, and the pushing accuracy and the pushing efficiency of the application programs are effectively improved.

In an embodiment, as shown in fig. 9, fig. 9 is a schematic flowchart of a step of training a prediction model, and specifically includes the following steps:

s902, acquiring historical application mirror image sets and historical application operation records of a plurality of users, and generating training sample data by using the historical application mirror image sets and the historical application operation records;

and S904, acquiring the retained application sets of a plurality of users, and generating the training labels by utilizing the retained application sets.

Before pushing the application program, the server needs to train a prediction model in advance, and specifically, the prediction model may be trained in advance in an offline training manner. The historical application mirror image set represents a set of all application programs installed in the user terminal at a certain designated time in the history, and the historical application operation record represents an application operation record of the user within a certain time period in the history. The reserved application set represents a set of all applications installed in the user terminal after a certain time period of a certain specified time of the history. For example, taking the designated time as T as an example, the historical application mirror image set and the historical application operation record are an application set and an application operation record in the user terminal at the time of T-1, the application set is kept as the set of the application programs in the user terminal at the time of T-1, and the time length corresponding to T-1 may be a period of one week, one month, and the like.

The server can acquire a large number of historical application mirror image sets, historical application operation records and reserved application sets of users, training sample data are generated by using the historical application mirror image sets and the historical application operation records, and the training sample data are used for training a preset neural network. The server generates training labels by utilizing the reserved application sets of the users, and the training labels are used for performing parameter adjustment and other processing on each training result so as to further train and optimize the prediction model.

S906, extracting the application mirror image characteristics of the historical application mirror image set, and extracting the application operation record matrix of the historical application operation record.

And S908, training the prediction model by using the application mirror image characteristics and the application operation record matrix and the training labels.

After a large number of historical application mirror image sets and historical application operation records of users are obtained by the server, the historical application mirror image sets and the historical application operation records are input into a preset neural network model, and feature extraction is carried out on the historical application mirror image sets to obtain corresponding application mirror image features. And extracting the characteristics of the historical application operation records to obtain a corresponding application operation record matrix. Specifically, the server may use an application category of an application program retained in the user terminal after the user is in a future period of time as a prediction target, and train the neural network model through a preset function by using the application mirror image feature and the application operation record matrix.

And in the process of training the prediction model, the server carries out iterative training on the preset neural network model for multiple times by using the application mirror image characteristics and the application operation record matrix, and obtains a corresponding training result each time. And the server adjusts the parameters of the preset neural network model according to the training result by using the training labels, and continues to carry out iterative training.

FIG. 10 is an overall framework of the application push process in one embodiment, as shown in FIG. 10, including a model training phase and an online recommendation phase. Taking the designated time as T as an example, T-1 represents a history stage, that is, before the server performs online recommendation, the server needs to train by using the historical application image set, the historical application operation records and the training labels to obtain a prediction model. T +1 represents a prediction result obtained by processing using a prediction model, that is, a push result. The predicted result may represent a future application of interest to the user for the node at time T. The characteristics of the user requirements in different periods are extracted through the prediction model, and the interestingness score of the application program to be pushed is determined, so that the more comprehensive and accurate interestingness requirements of the user can be effectively obtained, the application program is pushed according to the interestingness requirements of the user, and the accuracy of pushing the application program is effectively improved.

In one embodiment, training the predictive model using the application mirroring features and the application operation record matrix and the training labels comprises: inputting the application mirror image characteristics and the application operation record matrix into a preset neural network model to obtain a training result; and adjusting parameters of the neural network model and continuing training based on the difference between the training result and the training label until the training condition is met, and finishing the training to obtain the required prediction model.

The training condition is a condition satisfying model training. The training condition may be that a preset number of iterations is reached, or that the classification performance index of the image classifier after the parameters are adjusted reaches a preset index.

Specifically, the server trains a preset neural network model by using the application mirror image features and the application operation record matrix each time, obtains a corresponding training result, and then compares the training result with the training labels to obtain the difference between the training result and the training labels. And the server further adjusts the parameters of the preset neural network model by taking the reduction of the difference as a target, and continues training. And if the training result of the neural network model after parameter adjustment does not meet the training condition, continuously adjusting the parameters of the neural network model by using the training label and continuously training. And ending the training until the training condition is met to obtain the required prediction model.

For example, the difference between the training result and the training label can be measured by a cost function, and a cross entropy loss function or a mean square error function can be selected as the cost function. The server can finish training when the value of the cost function is smaller than a preset value, so that the accuracy of the user in predicting the installation behavior of the application program is improved.

In this embodiment, the historical application mirror image sets and the historical application operation records of the multiple users are used as training sample data, and the retained application set is used as a training label. When the prediction model is trained, the neural network model is guided, adjusted and optimized through the training labels, so that the prediction precision of the user on the installation behavior of the application program can be effectively improved, the prediction accuracy of the model is effectively improved, and the pushing accuracy of the application program can be effectively improved.

It should be understood that although the steps in the flowcharts of fig. 2, 3, 5, 6, 7, 9 are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not performed in the exact order shown and described, and may be performed in other orders, unless explicitly stated otherwise. Moreover, at least some of the steps in fig. 2, 3, 5, 6, 7, and 9 may include multiple steps or multiple stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of performing the steps or stages is not necessarily sequential, but may be performed alternately or alternatingly with other steps or at least some of the other steps.

In one embodiment, as shown in fig. 11, there is provided an apparatus 1100 for pushing an application, which may be a part of a computer device using a software module or a hardware module, or a combination of the two, the apparatus specifically includes: an application obtaining module 1102, a data obtaining module 1104, an interest predicting module 1106, and an application pushing module 1108, wherein:

an application obtaining module 1102, configured to obtain an application program to be pushed, where the application program to be pushed includes an application category;

the data acquisition module 1104 is configured to determine a push range according to the application category, and acquire an application mirror image set and an application operation record of each user identifier corresponding to the push range;

an interest prediction module 1106, configured to input the application image set and the application operation record into a trained prediction model, extract a first period requirement characteristic of each user identifier through a first overlay network of the prediction model, and extract a second period requirement characteristic of each user identifier through a second overlay network of the prediction model; fusing the first period demand characteristics and the second period demand characteristics through a characteristic connection layer of the prediction model to obtain target demand characteristics; determining the interestingness score of each user identifier corresponding to the application program to be pushed according to the target demand characteristics;

the application pushing module 1108 is configured to filter, according to the interestingness score, a target user identifier that meets a pushing condition, and push an application program to be pushed to a user terminal corresponding to the target user identifier.

In one embodiment, the interest prediction module 1106 is further configured to extract a first time period requirement characteristic of each user identifier according to the application image set by using a first overlay network of a prediction model; and extracting second-period demand characteristics of each user identification according to the application operation records by using a second overlay network of the prediction model.

In one embodiment, the interest prediction module 1106 is further configured to input the application image set to the first overlay network, perform feature extraction on the application installation images in the application image set, and obtain corresponding application image features; coding the application mirror image characteristics by using the first overlay network and the corresponding weight parameters to obtain coded application mirror image characteristics; and extracting the first time period demand characteristics of each user identifier according to the coded application mirror image characteristics by using a preset function.

In one embodiment, the application operation record comprises at least one of application identification, operation type and operation time; the operation type includes at least one of an application install operation and an application uninstall operation.

In one embodiment, the interest prediction module 1106 is further configured to input the application operation record into the second overlay network, perform feature extraction on the application operation record, and obtain an application embedding representation, an operation time representation, and an operation type representation; coding the application embedding representation, the application operation representation and the operation time representation by utilizing the second overlay network and corresponding weight parameters to obtain an application operation recording matrix; and extracting the second-period requirement characteristics of each user identifier according to the application operation recording matrix.

In one embodiment, the interest prediction module 1106 is further configured to obtain application image characteristics corresponding to the application image set; inputting the application mirror image characteristics into an application operation record matrix for characteristic fusion to obtain a fused application matrix; and extracting the second-period demand characteristics of each user identifier according to the application matrix.

In one embodiment, the interest prediction module 1106 is further configured to input the first period requirement characteristics and the second period requirement characteristics into a characteristic connection layer of the prediction model, splice the first period requirement characteristics and the second period requirement characteristics, and generate target requirement characteristics of each user identifier by using the spliced requirement characteristics.

In an embodiment, the application pushing module 1108 is further configured to calculate, according to the fused target demand characteristics, an interestingness score of an application category of the application program to be pushed corresponding to each user identifier; generating a user push list corresponding to each application category according to the interestingness score; and pushing the application programs to be pushed of each application category to the corresponding user terminals according to the user pushing list.

In an embodiment, as shown in fig. 12, the apparatus further includes a request obtaining module 1103, configured to receive an application obtaining request sent by a user terminal, where the application obtaining request carries a user identifier; the interest prediction module 1106 is further configured to obtain a current application image set and an application operation record of the user identifier, and input the current application image set and the application operation record into the prediction model to obtain an interest level score of each application category corresponding to the user identifier; the application pushing module 1108 is further configured to extract an application category meeting the condition according to the interestingness score, and obtain an application program to be pushed corresponding to the application category; and pushing the application program to be pushed to the user terminal.

In an embodiment, the application pushing module 1108 is further configured to obtain a heat value of each to-be-pushed application program corresponding to the application category; extracting an application program to be pushed, of which the heat value meets a condition threshold value; generating an application push list by the obtained application program to be pushed according to the interestingness score; and pushing the application program to be pushed in the application pushing list to the user terminal.

In one embodiment, as shown in fig. 13, the apparatus further includes a model training module 1101, configured to obtain historical application image sets and historical application operation records of multiple users, and generate training sample data using the historical application image sets and the historical application operation records; acquiring a reserved application set of a plurality of users, and generating a training label by using the reserved application set; extracting application mirror image characteristics of a historical application mirror image set, and extracting an application operation record matrix of a historical application operation record; the predictive model is trained using the application mirroring features and the application operation record matrix and the training labels.

In one embodiment, the model training module 1101 is further configured to input the application mirror image feature and the application operation record matrix to a preset neural network model to obtain a training result; and adjusting parameters of the neural network model and continuing training based on the difference between the training result and the training label until the training condition is met, and finishing the training to obtain the required prediction model.

For specific limitations of the apparatus for pushing the application, reference may be made to the above limitations of the method for pushing the application, which are not described herein again. The modules in the device for pushing the application program can be wholly or partially realized by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as shown in fig. 14. The computer device includes a processor, a memory, and a network interface connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The database of the computer equipment is used for storing data such as application programs to be pushed, application mirror image sets, application operation records and the like. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a method of pushing an application.

Those skilled in the art will appreciate that the architecture shown in fig. 14 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

In one embodiment, a computer device is further provided, which includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method embodiments when executing the computer program.

In an embodiment, a computer-readable storage medium is provided, in which a computer program is stored which, when being executed by a processor, carries out the steps of the above-mentioned method embodiments.

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 hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, storage, database or other medium used in the embodiments provided herein can include at least one of non-volatile and volatile memory. Non-volatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical storage, or the like. Volatile Memory can include Random Access Memory (RAM) or external cache Memory. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM), among others.

The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. A method of pushing an application, the method comprising:

inputting the application mirror image set and the application operation record into a trained prediction model, extracting first period demand characteristics of each user identifier according to the application mirror image set through a first overlay network of the prediction model, and extracting second period demand characteristics of each user identifier according to the application operation record through a second overlay network of the prediction model;

2. The method of claim 1, wherein the extracting, by the first overlay network of the predictive model, the first time period requirement characteristics of each subscriber identity from the set of application images comprises:

inputting the application mirror image set into a first overlay network, and performing feature extraction on an application installation mirror image in the application mirror image set to obtain corresponding application mirror image features;

encoding the application mirror image characteristics by using the first overlay network and corresponding weight parameters to obtain encoded application mirror image characteristics;

and extracting the first time period demand characteristics of each user identifier according to the coded application mirror image characteristics by using a preset function.

3. The method of claim 1, wherein the application operation record comprises an application identification, an operation type, an operation time; the operation type includes at least one of an application install operation and an application uninstall operation.

4. The method of claim 3, wherein said extracting second age requirement characteristics for respective user identities from said application operational record via said second overlay network of said predictive model comprises:

inputting the application operation record into a second overlay network, and performing feature extraction on the application operation record to obtain an application embedded representation, an operation time representation and an operation type representation;

encoding the application embedding representation, the application operation representation and the operation time representation by using the second overlay network and corresponding weight parameters to obtain an application operation recording matrix;

and extracting the second period requirement characteristics of each user identifier according to the application operation recording matrix.

5. The method according to claim 4, wherein the extracting the second time period requirement characteristics of each user identifier according to the application operation record matrix comprises:

acquiring application mirror image characteristics corresponding to the application mirror image set;

inputting the application mirror image characteristics into the application operation record matrix for characteristic fusion to obtain a fused application matrix; and extracting the second period demand characteristics of each user identifier according to the application matrix.

6. The method of claim 1, wherein said fusing the first and second term demand characteristics through a feature connection layer of the predictive model comprises:

inputting the first and second term demand characteristics into a characteristic connection layer of the predictive model;

and splicing the first period demand characteristics and the second period demand characteristics, and generating target demand characteristics of each user identifier by using the spliced demand characteristics.

7. The method of any one of claims 1 to 6, further comprising:

determining the interestingness score of each user identification corresponding to the application category of the application program to be pushed according to the target demand characteristics;

generating a user push list corresponding to each application category according to the interestingness score;

and pushing the application programs to be pushed of each application category to corresponding user terminals according to the user pushing list.

8. The method of claim 1, further comprising:

receiving an application acquisition request sent by a user terminal, wherein the application acquisition request carries a user identifier;

acquiring a current application mirror image set and an application operation record of the user identification, and inputting the current application mirror image set and the application operation record into the prediction model to obtain interestingness scores of the user identification corresponding to each application category;

screening application categories meeting conditions according to the interestingness scores, and acquiring application programs to be pushed corresponding to the application categories;

and pushing the application program to be pushed to the user terminal.

9. The method of claim 8, further comprising:

acquiring the heat value of each application program to be pushed corresponding to the application type;

extracting the application program to be pushed, of which the heat value meets a condition threshold value;

generating an application pushing list for the obtained application program to be pushed according to the interestingness score;

and pushing the application program to be pushed in the application pushing list to the user terminal.

10. The method of claim 1, wherein the step of training the predictive model comprises:

acquiring historical application mirror image sets and historical application operation records of a plurality of users, and generating training sample data by using the historical application mirror image sets and the historical application operation records;

acquiring a retention application set of a plurality of users, and generating a training label by using the retention application set;

extracting application mirror image characteristics of the historical application mirror image set, and extracting an application operation record matrix of the historical application operation record;

and training a prediction model by using the application mirror image characteristics, the application operation record matrix and the training labels.

11. The method of claim 10, wherein the training a predictive model using the application mirror features and the application operation record matrix and the training labels comprises:

inputting the application mirror image characteristics and the application operation recording matrix into a preset neural network model to obtain a training result;

and adjusting parameters of the neural network model and continuing training based on the difference between the training result and the training label until the training condition is met, and finishing the training to obtain the required prediction model.

12. An apparatus for pushing an application, the apparatus comprising:

the interest prediction module is used for inputting the application mirror image set and the application operation record into a trained prediction model, extracting first period demand characteristics of each user identifier according to the application mirror image set through a first overlay network of the prediction model, and extracting second period demand characteristics of each user identifier according to the application operation record through a second overlay network of the prediction model; fusing the first period demand characteristics and the second period demand characteristics through a characteristic connection layer of the prediction model to obtain target demand characteristics; determining the interestingness score of each user identification corresponding to the application program to be pushed according to the target demand characteristics;

13. The apparatus according to claim 12, further comprising a model training module for obtaining a set of historical application images and a historical application operation record of a plurality of users, and generating training sample data by using the set of historical application images and the historical application operation record; acquiring a retention application set of a plurality of users, and generating a training label by using the retention application set; extracting application mirror image characteristics of the historical application mirror image set, and extracting an application operation record matrix of the historical application operation record; and training a prediction model by using the application mirror image characteristics, the application operation record matrix and the training labels.

14. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor realizes the steps of the method of any one of claims 1 to 11 when executing the computer program.

15. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 11.