Content pushing method and device, and storage medium

Abstract

Embodiments of the present application relate to the technical field of computer, and disclose a content pushing method and device and a storage medium, the method comprising: obtaining a plurality of user features of a target user, the plurality of user features of the target user being user features of the target user in different dimensions; performing feature fusion processing on the plurality of user features of the target user through a trained content type prediction model to obtain a type representation vector of the target user, and performing prediction processing on a content type of interest of the target user based on the type representation vector of the target user to obtain a target content type of interest of the target user; and determining a target to-be-pushed content from a plurality of to-be-pushed contents, the target to-be-pushed content matching the target content type of interest of the target user, and the target to-be-pushed content being used for pushing to the target user. By using the embodiments of the present application, personalized determination of to-be-pushed content can be achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computers, and particularly relates to a content pushing method and device and a storage medium. BACKGROUND

[0002] In a related scenario of content pushing, an application can push content related to a service provided by the application to a user. For example, if the application is a news application, the application can push various news content to the user, and if the application is a video application, the application can push various video content to the user. When various applications push content, personalized pushing for different users is pursued to improve user experience, and thus improve user activity and user stickiness. However, in the existing personalized content pushing method for different users, manual selection by business personnel is relied on, which leads to homogenization and weak matching of pushed content, that is, the personalized pushed content for a user cannot meet the interest preferences of the user, and thus leads to problems such as reduced user experience, user activity and user stickiness. SUMMARY

[0003] Embodiments of the present application provide a content pushing method, device, equipment and storage medium, which can realize personalized determination of to-be-pushed content.

[0004] In one aspect, the present application provides a content pushing method, comprising:

[0005] obtaining a plurality of user features of a target user, the plurality of user features of the target user being user features of the target user in different dimensions;

[0006] performing feature fusion processing on the plurality of user features of the target user through the trained content type prediction model to obtain a type representation vector of the target user, and performing prediction processing on a content type concerned by the target user based on the type representation vector of the target user to predict a target content type concerned by the target user;

[0007] determining a target to-be-pushed content from a plurality of to-be-pushed contents, the target to-be-pushed content matching the target content type concerned by the target user, and the target to-be-pushed content being used for pushing to the target user.

[0008] In one aspect, the present application provides a content pushing device, comprising an obtaining unit and a processing unit, and the obtaining unit and the processing unit are used to execute the above-mentioned content pushing method.

[0009] In one aspect, the present application provides a content pushing equipment, characterized in that the content pushing equipment comprises an input interface and an output interface, and further comprises:

[0010] a processor adapted to implement one or more instructions; and

[0011] a computer storage medium storing one or more instructions adapted to be loaded and executed by the processor to implement the content pushing method.

[0012] In an aspect, an embodiment of the present application provides a computer storage medium, characterized in that the computer storage medium stores computer program instructions, and the computer program instructions are executed by a processor to implement the content pushing method.

[0013] In an aspect, an embodiment of the present application provides a computer program product or a computer program, the computer program product or the computer program comprising computer instructions stored in a computer readable storage medium; a processor of a content pushing device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, and the computer instructions are executed by the processor to implement the content pushing method.

[0014] In an embodiment of the present application, the content pushing device can obtain a plurality of user features of a target user, the plurality of user features of the target user being user features of the target user in different dimensions; the plurality of user features of the target user are processed by feature fusion through a trained content type prediction model to obtain a type representation vector of the target user, and a content type that the target user pays attention to is predicted based on the type representation vector of the target user, and a target content type that the target user pays attention to is predicted; a target to-be-pushed content is determined from a plurality of to-be-pushed contents, the target to-be-pushed content matches the target content type that the target user pays attention to, and the target to-be-pushed content is used to push to the target user. The content pushing device can predict the target content type that the user pays attention to according to the plurality of user features of the user, and then determine the target to-be-pushed content matching the target content type that the user pays attention to from the to-be-pushed contents, that is, the personalized prediction of the content type that the user pays attention to can be realized for different users, and then the target to-be-pushed content is determined from the to-be-pushed contents according to the target content type that the user pays attention to predicted, the personalized determination of the to-be-pushed content can be realized, and then the personalized pushing of the content can be realized to improve the user experience, the user activity, and the user stickiness. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0016] Figure 1 is a structural schematic diagram of a content pushing system provided by an embodiment of the present application;

[0017] Figure 2 is a training flow schematic diagram of a content type prediction model provided by an embodiment of the present application;

[0018] Figure 3 is a schematic diagram of a content detail page provided by an embodiment of the present application;

[0019] Figure 4 is a flow schematic diagram of training a content type prediction model provided by an embodiment of the present application;

[0020] Figure 5 is a schematic diagram of a content type that a training user is interested in, which is obtained by a content type prediction model provided by an embodiment of the present application;

[0021] Figure 6 is a training flow schematic diagram of another content type prediction model provided by an embodiment of the present application;

[0022] Figure 7 is a schematic diagram of a content type that a training user is interested in, which is obtained by another content type prediction model provided by an embodiment of the present application;

[0023] Figure 8 is a training flow schematic diagram of another content type prediction model provided by an embodiment of the present application;

[0024] Figure 9a is a schematic diagram of a daily report type content card provided by an embodiment of the present application;

[0025] Figure 9b is a schematic diagram of another daily report type content card provided by an embodiment of the present application;

[0026] Figure 9c is a schematic diagram of a hot spot type content card provided by an embodiment of the present application;

[0027] Figure 9d is a schematic diagram of a social type content card provided by an embodiment of the present application;

[0028] Figure 10is a schematic diagram of training user's multiple user features provided by an embodiment of the present application;

[0029] Figure 11 is another schematic diagram of predicting the content type that the training user is interested in through a content type prediction model provided by an embodiment of the present application;

[0030] Figure 12 is another schematic diagram of predicting the content type that the training user is interested in through a content type prediction model provided by an embodiment of the present application;

[0031] Figure 13 is a flowchart of a content pushing method provided by an embodiment of the present application;

[0032] Figure 14 is a schematic diagram of obtaining the target content type that the target user is interested in through a trained content type prediction model provided by an embodiment of the present application;

[0033] Figure 15 is another schematic diagram of obtaining the target content type that the target user is interested in through a trained content type prediction model provided by an embodiment of the present application;

[0034] Figure 16 is a structural schematic diagram of a content pushing device provided by an embodiment of the present application;

[0035] Figure 17 is a structural schematic diagram of a content pushing device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0037] In order to realize personalized determination of the to-be-pushed content, the present application provides a content pushing scheme. The content type that the target user is interested in can be predicted according to multiple user features of the target user through a trained content type prediction model. The target to-be-pushed content is determined from multiple to-be-pushed contents. The target to-be-pushed content matches the target content type that the target user is interested in. The target to-be-pushed content is used to push to the target user.

[0038] The above content pushing scheme can be executed by a content pushing device. The content pushing device can be a terminal device integrated with a content pushing service, such as a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart vehicle, a smart wearable device, or the like. The content pushing device can also be a server capable of providing a content pushing service. The server can be a physical server, a server cluster composed of multiple physical servers, or a distributed system. The server can also be a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and basic cloud computing services such as big data and artificial intelligence platforms.

[0039] Based on the above content pushing scheme, an embodiment of the present application provides a content pushing system. Referring to Figure 1 FIG. 1 is a structural schematic diagram of a content pushing system provided by an embodiment of the present application. Figure 1 The content pushing system shown in FIG. 1 can include a content pushing device 101 and a terminal device 102. The content pushing device 101 can be a server, such as a physical server, a server cluster composed of multiple physical servers, or a distributed system. The content pushing device 101 can also be a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and basic cloud computing services such as big data and artificial intelligence platforms. The terminal device 102 can include any one or more of a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart vehicle, and a smart wearable device. The content pushing device 101 and the terminal device 102 can be directly or indirectly connected through wired or wireless communication, which is not limited in the present application.

[0040] In an embodiment, the terminal device 102 can run a target application. The target application can be any application running in the terminal device 102 and capable of providing a content pushing service, such as a social application, a news application, or a music application. The content pushing device 101 is a content pushing device corresponding to the target application and provides service support for the target application, specifically, provides content pushing service support for the target application. For example, if the target application is a news application, the content pushing device 101 is a content pushing device corresponding to the news application and can provide content pushing service support for the news application.

[0041] In one embodiment, the content push device 101 can predict the target content type that the target user is interested in based on multiple user characteristics of the target user through a trained content type prediction model; determine the target content to be pushed among multiple content to be pushed, the target content to be pushed matches the target content type that the target user is interested in, and the target content to be pushed is used to push to the target user, for example, the content push device 101 can push the target content to be pushed to the target user's terminal device 102.

[0042] The trained content type prediction model is obtained by training multiple user features of the training users and the labeled content types that the training users are interested in. Based on this, this application embodiment introduces the training process of the content type prediction model, such as... Figure 2 The diagram shown is a schematic representation of the training process of a content type prediction model provided in an embodiment of this application. Figure 2 The training process of the content type prediction model shown can be executed by a content delivery device or by any other electronic device capable of implementing the training process of the content type prediction model. This application embodiment uses a content delivery device for description. Figure 2 The training process for the content type prediction model shown may include the following steps:

[0043] S201, Obtain training samples.

[0044] The training samples include multiple user features of the training users and labeled content types that the training users have followed. The multiple user features represent the user characteristics of the training users across different dimensions; different dimensions of user features can be obtained according to different training needs. In other words, the user features obtained will differ depending on the specific training requirements. The labeled content types that the training users have followed represent the types of historical push notifications that the training users have viewed on the content details page, and can characterize the training users' interests and preferences.

[0045] For example, if the training requirement is that the trained content type prediction model can accurately predict the target content type that the target user is interested in, where the target content type is included in the preset content type, that is, the trained content type prediction model can accurately predict the target content type that the target user is interested in from the preset content type. Where the preset content type is set in advance according to business requirements, it is the content type of the content that the content push service provided by the target application can push. For example, if the target application is a video application, and the preset content type is set to movie type, TV series type, variety show type, and documentary type according to business requirements, the trained content type prediction model is required to be able to predict the target content type that the target user is interested in from the movie type, TV series type, variety show type, and documentary type based on the multiple user features of the target user; if the preset content type is set to long video type and short video type according to business requirements, the trained content type prediction model is required to be able to predict the target content type that the target user is interested in from the long video type and short video type based on the multiple user features of the target user. Further, based on different training requirements, that is, based on different preset content types, the multiple user features of the training user may be different, and the preset content type can be set according to business requirements and the multiple user features of the training user used when designing the training content type prediction model can be designed according to business requirements.

[0046] In a specific implementation, the content push device obtaining the training sample can include: determining historical push content pushed to the user in a preset time period and the user browsing the content detail page of the historical push content; determining the user as a training user and obtaining multiple user features of the training user; labeling the content type of the historical push content as the content type that the training user is interested in; and constructing the training sample based on the multiple user features of the training user and the labeled content type that the training user is interested in. Where the preset time period is set according to specific requirements, for example, the time point corresponding to the multiple user features of the training user obtained by the content push device can be determined as the end time point of the preset time period, and the length of the preset time period can be set according to requirements. For example, if the length of the preset time period is 1 hour and the time point corresponding to the multiple user features of the training user obtained by the content push device is 16:00, the historical push content pushed to the user in the preset time period is the historical push content pushed to the user from 15:00 to 16:00.

[0047] For example, if there are user 1, user 2, user 3, user 4 and user 5, the historical push content pushed to the users in a preset time period is content 1, content 2 and content 3; if user 1 and user 2 have browsed the content detail page of content 1, user 3 has browsed the content detail page of content 2, and no user has browsed the content detail page of content 3, user 1 and user 2 are determined as training users, and the content type of content 1 is marked as the content type that the training user 1 (user 1) and the training user 2 (user 2) are interested in; user 3 is determined as a training user, and the content type of content 2 is marked as the content type that the training user 3 (user 3) is interested in. If the content type of content 1 is type 1, and the content type of content 2 is type 2, the marked content type that the training user 1 is interested in is type 1, the marked content type that the training user 2 is interested in is type 1, and the marked content type that the training user 3 is interested in is type 2. The user who has browsed the content detail page of the historical push content is determined as a training user, and the content type of the historical push content browsed by the training user is marked as the content type that the training user is interested in, which can ensure the accuracy of the marked content type that the training user is interested in, so that the trained content type prediction model can accurately predict the content type that the user is interested in.

[0048] In one embodiment, the marked content type that the training user is interested in can be represented in the form of a vector based on the preset content type. For example, if the preset content types are type 1, type 2 and type 3, the corresponding relationship between the vector element and the preset content type is {type 1, type 2, type 3}, and if the marked content type that the training user is interested in is type 1, the marked content type that the training user is interested in can be represented in the form of a vector as {1, 0, 0} based on the preset content type.

[0049] In one embodiment, if the target application is a social application that provides a content push service, the content detail page of the historical push content can be marked as shown in Figure 3 In one embodiment, if the target application is a social application that provides a content push service, the content detail page of the historical push content can be marked as shown in Figure 3As shown, a schematic diagram of a content detail page provided by an embodiment of the present application is shown, taking user 1 as an example. If user 1 triggers the content push function entry as indicated by 301, the content push main page as indicated by 302 is entered, that is, the content push main page as indicated by 302 is displayed in the terminal device of user 1. The content push main page includes the title page of the historical push content pushed to user 1. Optionally, the title page of the historical push content can be carried in the form of a card. If the title pages of the historical push content pushed to user 1 included in the content push main page are the title page of content 1, the title page of content 2, and the title page of content 3, if user 1 triggers the title page of content 1, the content detail page of content 1 as indicated by 303 is displayed in the terminal device of user 1.

[0050] In an embodiment, the training user can be determined by performing a point on the content detail page of the historical push content, and then the plurality of user features of the training user are obtained. Since the plurality of user features of the training user obtained by different training requirements are different, taking the plurality of user features of the training user as the first user feature of the training user and the second user feature of the training user as an example, since the original feature data of the first user feature of the training user and the second user feature of the training user obtained by the content push device may have data missing and the like, after the content push device obtains the original feature data, the original feature data needs to be preprocessed, for example, missing value filling and the like, to obtain preprocessed feature data that can be used as model input of the content type prediction model. It should be noted that the feature data involved in the embodiments of the present application are all preprocessed feature data. As shown in Figure 4 As shown, a schematic diagram of a content type prediction model training process provided by an embodiment of the present application is shown. The content push device obtains the original feature data of the plurality of user features of the training user; the original feature data is preprocessed to obtain preprocessed feature data; the training sample is constructed based on the preprocessed feature data; the content type prediction model is trained based on the training sample to obtain the trained content type prediction model.

[0051] S202, the plurality of user features of the training user are used as model input, the content type focused on by the training user is labeled as expected output, the content type prediction model is trained based on the training sample, and the trained content type prediction model is obtained.

[0052] Among them, the plurality of user features of the training user are user features of the training user in different dimensions; the content type focused on by the training user is the content type in the preset content type.

[0053] In one embodiment, the content push device trains a content type prediction model based on training samples to obtain a trained content type prediction model. This process may include: using the content type prediction model to perform feature fusion processing on multiple user features of the training user to obtain a type representation vector for the training user; predicting the content types followed by the training user based on the type representation vector to obtain the predicted content types followed by the training user; and training the content type prediction model based on the predicted content types followed by the training user and the labeled content types followed by the training user to obtain a trained content prediction model.

[0054] In one embodiment, the content push device uses a content type prediction model to perform feature fusion processing on multiple user features of a training user to obtain a type representation vector for the training user. This is achieved by processing the feature representation vectors of the multiple user features of the training user. Specifically, the process of using the content push device to perform feature fusion processing on multiple user features of a training user to obtain a type representation vector may include: performing feature embedding processing on each of the multiple user features of the training user to obtain multiple feature representation vectors for the training user, wherein different user features of the training user correspond to different feature representation vectors for the training user; and performing feature fusion processing on the multiple feature representation vectors of the training user to obtain the type representation vector for the training user.

[0055] like Figure 5 The diagram illustrates an embodiment of this application that uses a content type prediction model to predict the content types a training user is interested in. The content type prediction model includes an input layer, an embedding layer, a network mapping layer, and an output layer. The content push device uses the content type prediction model to perform feature fusion processing on multiple user features of the training user to obtain a type representation vector for the training user. This process includes: receiving multiple user features of the training user through the input layer and inputting these features into the embedding layer; performing feature embedding processing on each of the multiple user features of the training user through the embedding layer to obtain multiple feature representation vectors for the training user, where different user features correspond to different feature representation vectors for the training user; and performing feature fusion processing on the multiple feature representation vectors of the training user through the network mapping layer to obtain a type representation vector for the training user. Further, the output layer predicts the content types the training user is interested in based on the type representation vectors of the training user, thus obtaining the predicted content types the training user is interested in.

[0056] For example, if the one user feature includes multiple sub-features, when the one user feature of the training user is processed by the embedding layer for feature embedding, the embedding layer first obtains a low-dimensional feature representation vector of each sub-feature in the one user feature, and splices the obtained low-dimensional feature representation vector of each sub-feature to obtain a training user one feature representation vector corresponding to the one user feature, wherein the training user one feature representation vector is a high-dimensional feature representation vector. For example, if the one user feature includes feature 1, feature 2 and feature 3, the low-dimensional feature representation vector of feature 1 obtained by the embedding layer is {vector 1}, the low-dimensional feature representation vector of feature 2 is {vector 2}, and the low-dimensional feature representation vector of feature 3 is {vector 3}, and the training user one feature representation vector corresponding to the one user feature can be {vector 1, vector 2, vector 3}.

[0057] Further, when the content type prediction model is trained according to different training requirements, the number of user features of the training user may be different, and thus the number of feature representation vectors of the training user obtained may be different. Due to the difference in the number of feature representation vectors of the training user, the specific process of processing the multiple feature representation vectors of the training user by the network mapping layer for feature fusion to obtain the type representation vector of the training user is different. The specific process of processing the multiple feature representation vectors of the training user for feature fusion to obtain the type representation vector of the training user will be introduced in subsequent embodiments.

[0058] Further, the content pushing device predicts the content type that the training user pays attention to based on the type feature vector of the training user through an output layer, predicts the content type that the training user pays attention to, and obtains the predicted content type that the training user pays attention to. The output layer can be a normalized exponential output layer, that is, a softmax layer. The type feature vector of the training user is processed through the softmax layer to obtain the predicted content type that the training user pays attention to. In a specific implementation, the type feature vector of the training user can be mapped to a probability value of a preset content type through the softmax layer. The predicted content type that the training user pays attention to is the preset content type corresponding to the maximum probability value. For example, if the preset content types are type 1, type 2, and type 3, the type feature vector of the training user is mapped to probability values of type 1, type 2, and type 3, which are 99%, 0.5%, and 0.5% respectively, and the predicted content type that the training user pays attention to is type 1. Optionally, the predicted content type that the training user pays attention to can be represented in the form of a vector, and the vector elements are preset content types. If the correspondence between the vector elements and the preset content types is {type 1, type 2, type 3}, the predicted content type that the training user pays attention to can be represented as {99%, 0.5%, 0.5%}.

[0059] Further, the content pushing device trains the content type prediction model based on the predicted content type that the training user pays attention to and the labeled content type that the training user pays attention to, and obtains the trained content prediction model, which can include: determining the loss value of the loss function based on the predicted content type that the training user pays attention to and the labeled content type that the training user pays attention to, training the content type prediction model based on the loss value generated by different training samples, so that the loss value generated by the test sample used to test the content type prediction model is less than the preset loss value, thereby obtaining the trained content type prediction model. Wherein, the test sample is obtained in the same way as the training sample, which will not be repeated here, and the test sample and the training sample are different samples; the preset loss value is set according to the training requirement, which is the training stop condition of the content type prediction model. Specifically, the content pushing device can iteratively update the model parameters of the content type prediction model based on the loss value generated by different training samples, so that the loss value generated by the test sample used to test the content type prediction model based on the updated model parameters is less than the preset loss value, and then the trained content type prediction model is obtained based on the updated model parameters. Optionally, the model parameters of the content type prediction model can be iteratively updated based on the Adam optimization algorithm in the optimization algorithm.

[0060] In an embodiment, the loss function can be a multi-class cross-entropy loss function. The multi-class cross-entropy loss function can be determined by formula (1):

[0061]

[0062] wherein n is the number of preset content types, i is an integer variable, m i is a feature value of the i-th preset content type in the annotated content type of interest of the training user represented in a vector form, is a feature value of the i-th preset content type in the predicted content type of interest of the training user represented in a vector form, i.e., the probability of predicting the i-th preset content type. For example, if the preset content types are type 1, type 2 and type 3, the correspondence between the preset content types and the vector elements is {type 1, type 2, type 3}, if the annotated content type of interest of the training user represented in a vector form is {1, 0, 0}, and if the predicted content type of interest of the training user represented in a vector form is {99%, 0.5%, 0.5%}, the loss value of the multi-class cross-entropy loss function is: loss = -(1 x log(99%) + 0 x log(0.5%) + 0 x log(0.5%)).

[0063] In the embodiments of the present application, the content pushing device determines a user who has browsed the content detail page of the historical pushing content pushed to the user in a preset time period as a training user, obtains a plurality of user features of the training user, and determines the content type of the historical pushing content browsed by the training user in the preset time period as the annotated content type of interest of the training user. Based on the plurality of user features of the training user and the annotated content type of interest of the training user, a training sample is constructed. The plurality of user features of the training user are taken as the model input, and the annotated content type of interest of the training user is taken as the expected output. The content type prediction model is trained based on the training sample, and a trained content type prediction model is obtained. The trained content type prediction model can accurately predict the content type of interest of a target user based on a plurality of user features of the target user, i.e., can individually predict the content type of interest of different users based on a plurality of user features of the different users.

[0064] When the content type prediction model is trained based on the plurality of user features of the training user, the number of user features of the training user used may be different, so the process of performing feature fusion processing on the plurality of user features of the training user by the content type prediction model to obtain the type representation vector of the training user may be different. Based on this, the embodiments of the present application provide another training process of the content type prediction model. As shown in FIG. 6, the process of training the content type prediction model based on the plurality of user features of the training user includes the following steps. Figure 6As shown, a schematic diagram of a training process of another content type prediction model provided by the application embodiment is shown.

[0065] Figure 6 The training process of the content type prediction model shown can be performed by the content pushing device, or by any electronic device capable of implementing the training process of the content type prediction model. The application embodiment is introduced by the content pushing device. Figure 6 The training process of the content type prediction model shown can include the following steps:

[0066] S601, obtaining a training sample.

[0067] The training sample includes a plurality of user features of a training user and a labeled content type of which the training user is interested. The plurality of user features of the training user are user features of the training user in different dimensions. The labeled content type of which the training user is interested is a content type of a historical push content of which the training user has browsed the content detail page, and can represent the interest preference of the training user.

[0068] S602, performing feature fusion processing on the plurality of user features of the training user by the content type prediction model to obtain a type representation vector of the training user.

[0069] In a specific implementation, the content pushing device performs feature fusion processing on the plurality of user features of the training user by the content type prediction model to obtain the type representation vector of the training user, which can include: performing feature coupling processing on the plurality of user features of the training user by at least one coupler in the content type prediction model to obtain at least one coupled representation vector of the training user; and performing feature analysis processing on the at least one coupled representation vector of the training user to obtain the type representation vector of the training user.

[0070] The content type prediction model includes at least one coupler, and each coupler in the at least one coupler processes a certain number of user features of the training user, and the user features processed by different couplers are not completely the same. Optionally, the number of couplers in the content type prediction model and the certain number of user features of the training user that need to be coupled and processed by the same coupler can be set according to training requirements. For example, if 2 user features of the training user are used, the number of couplers in the content type prediction model and the certain number of user features of the training user that need to be coupled and processed by the same coupler can be set according to training requirements, for example, the number of couplers in the content type prediction model can be set to 1, and the 2 user features of the training user can be coupled and processed by the coupler; for another example, if 4 user features of the training user are used, the number of couplers in the content type prediction model and the certain number of user features of the training user that need to be coupled and processed by the same coupler can be set according to training requirements, for example, each two different user features of the 4 user features of the training user can be coupled and processed by different couplers; each three different user features of the 4 user features of the training user can be coupled and processed by different couplers; all the user features of the 4 user features of the training user can be coupled and processed by the same coupler, and the like.

[0071] In an embodiment, the content pushing device couples and processes the multiple user features of the training user through at least one coupler in the content type prediction model to obtain at least one coupled feature representation vector of the training user, which is obtained by processing a feature representation vector of the multiple user features of the training user. That is, before the content pushing device couples and processes the multiple user features of the training user through at least one coupler in the content type prediction model to obtain at least one coupled feature representation vector of the training user, the content pushing device also respectively processes each user feature of the multiple user features of the training user through the content type prediction model to obtain multiple feature representation vectors of the training user, wherein different user features of the training user correspond to different feature representation vectors of the training user. Further, the content pushing device couples and processes the multiple user features of the training user through at least one coupler in the content type prediction model to obtain at least one coupled feature representation vector of the training user can include: coupling and processing the multiple feature representation vectors of the training user through at least one coupler in the content type prediction model to obtain at least one coupled feature representation vector of the training user.

[0072] As Figure 7As shown, another schematic diagram of predicting the content type that the training user pays attention to through the content type prediction model provided by the embodiment of the present application is shown, and the content type prediction model includes an input layer, an embedding layer, a network mapping layer, and an output layer. The network mapping layer includes at least one coupler and a fully connected layer. Assuming that four user features of the training user are used according to the training requirement, which are the first user feature of the training user, the second user feature of the training user, the third user feature of the training user, and the fourth user feature of the training user, and the second user feature of the training user, the third user feature of the training user, and the fourth user feature of the training user need to be coupled with the first user feature of the training user for feature coupling processing, respectively, the content pushing device receives the first user feature of the training user, the second user feature of the training user, the third user feature of the training user, and the fourth user feature of the training user through the input layer, and inputs the first user feature of the training user, the second user feature of the training user, the third user feature of the training user, and the fourth user feature of the training user to the embedding layer through the input layer; the first user feature of the training user, the second user feature of the training user, the third user feature of the training user, and the fourth user feature of the training user are processed for feature embedding through the embedding layer, respectively, to obtain a plurality of feature representation vectors of the training user (the first user feature representation vector of the training user, the second user feature representation vector of the training user, the third user feature representation vector of the training user, and the fourth user feature representation vector of the training user, respectively); the first user feature representation vector of the training user and the second user feature representation vector of the training user are processed for feature coupling through a coupler (coupler 1) in the network mapping layer, to obtain a coupled representation vector of the training user (coupled representation vector 1 of the training user); the first user feature representation vector of the training user and the third user feature representation vector of the training user are processed for feature coupling through a coupler (coupler 2) in the network mapping layer, to obtain a coupled representation vector of the training user (coupled representation vector 2 of the training user); the first user feature representation vector of the training user and the fourth user feature representation vector of the training user are processed for feature coupling through a coupler (coupler 3) in the network mapping layer, to obtain a coupled representation vector of the training user (coupled representation vector 3 of the training user); the plurality of coupled representation vectors of the training user obtained based on different couplers are processed for feature analysis through the fully connected layer in the network mapping layer, to obtain a type representation vector of the training user; the content type that the training user pays attention to is predicted through the output layer based on the type representation vector of the training user, to obtain the predicted content type that the training user pays attention to.

[0073] In an embodiment, the content pushing device performs feature coupling processing on the multiple feature representation vectors of the training user through a coupler in the content type prediction model to obtain a coupled representation vector of the training user. Before this, the multiple feature representation vectors of the training user to be processed by the coupler need to be spliced to obtain a spliced feature representation vector to be processed by the coupler. The spliced feature representation vector to be processed by the coupler is input into the coupler, so that the coupler performs feature coupling processing on the spliced feature representation vector to be processed by the coupler to realize feature coupling processing on the multiple feature representation vectors of the training user. For example, if the multiple feature representation vectors of the training user to be processed by the coupler are {feature representation vector 1} and {feature representation vector 2}, the spliced feature representation vector to be processed by the coupler can be {feature representation vector 1, feature representation vector 2} or {feature representation vector 2, feature representation vector 1}.

[0074] In an embodiment, before the content pushing device performs feature analysis processing on the multiple coupled representation vectors of the training user obtained based on different couplers through a fully connected layer to obtain a type representation vector of the training user, the multiple coupled representation vectors of the training user are spliced to obtain a coupled spliced vector of the training user. The content pushing device can perform feature analysis processing on the multiple coupled representation vectors of the training user obtained based on different couplers through a fully connected layer to obtain a type representation vector of the training user, which can include performing feature analysis processing on the coupled spliced vector of the training user to obtain a type representation vector of the training user. For example, if a coupled representation vector of the training user is {coupled representation vector 1}, a coupled representation vector of the training user is {coupled representation vector 2}, and a coupled representation vector of the training user is {coupled representation vector 3}, the coupled spliced vector of the training user can be {coupled representation vector 1, coupled representation vector 2, coupled representation vector 3}.

[0075] S603, performing prediction processing on the content type concerned by the training user based on the type representation vector of the training user to obtain a predicted content type concerned by the training user.

[0076] S604, training the content type prediction model based on the predicted content type concerned by the training user and the labeled content type concerned by the training user to obtain a trained content prediction model.

[0077] In this embodiment, the content push device trains a content type prediction model based on multiple user features of the training user and the content types that the training user is interested in. This enables the trained content type prediction model to accurately predict the content types that the target user is interested in based on multiple user features of the target user. In other words, it can make personalized predictions of the content types that different users are interested in based on multiple user features of different users.

[0078] Based on the above system and method embodiments, this application provides another training process for a content type prediction model. For example... Figure 8 The diagram shown is a schematic diagram of the training process of another content type prediction model provided in an embodiment of this application. Figure 8 The training process of the content type prediction model shown uses multiple user features of the training users, including: user-side features of the training users, push-side features of the training users, social-side features of the training users, and contextual features of the training users. Figure 8 The training process of the content type prediction model shown can be executed by a content delivery device or by any other electronic device capable of implementing the training process of the content type prediction model. This application embodiment uses a content delivery device for description. Figure 8 The training process for the content type prediction model shown may include the following steps:

[0079] S801, Obtain training samples.

[0080] The training samples include multiple user features of the training users and labeled content types that the training users are interested in. The multiple user features represent the user characteristics of the training users across different dimensions. The labeled content types that the training users are interested in are the types of historical push content that the training users have viewed on the content details page, which can characterize the training users' interests and preferences.

[0081] In one embodiment, if the target application is a social application that provides content push services, and the preset content types are set as daily reports, trending topics, and social topics based on business needs, then the daily report content consists of content that users are interested in and prefer; the trending topics content consists of real-time trending events; and the social topics content consists of high-interaction, high-quality content within the user's social circle, such as content liked, commented on, or shared by the user's social friends. Each content type can be either text or video content. Optionally, the title page for different content types can be displayed in a card format, i.e., in the form of content cards, such as... Figure 9aAs shown in FIG. 9, a schematic diagram of a content card of a daily report type provided by an embodiment of the present application is shown. The content card of the daily report type can be specifically as shown by the label 901. The content is specifically a graphic-text content. As shown in FIG. 10, a schematic diagram of another content card of a daily report type provided by an embodiment of the present application is shown. The content card of the daily report type can be specifically as shown by the label 902. The content is specifically a video content. As shown in FIG. 11, a schematic diagram of a content card of a hot spot type provided by an embodiment of the present application is shown. The content card of the hot spot type can be specifically as shown by the label 903. As shown in FIG. 12, a schematic diagram of a content card of a social type provided by an embodiment of the present application is shown. The content card of the social type can be specifically as shown by the label 904. The content is specifically a content liked by a social friend of a user. Figure 9b Figure 9c Figure 9d

[0082] ​​​Further, if the multiple user features of the training user for the content type prediction model training are set based on the training requirement and are respectively a user side feature of the training user, a push side feature of the training user, a social side feature of the training user, and a context feature of the training user, wherein the user side feature of the training user is used to represent the basic feature of the training user, and can include gender feature, age feature, occupation feature, city feature, and click feature, etc., wherein the gender feature indicates that the gender of the training user is male or female; the age feature indicates the age of the training user, which can be represented by an integer value, for example, 12; the occupation feature indicates the occupation of the training user, for example, teacher, lawyer, etc.; the city feature indicates the city where the training user usually resides, for example, Beijing, Shanghai, etc.; and the click feature represents the click rate of the training user on the historical push content pushed to the training user within a first preset time period, which can be represented by a floating point number value, for example, 0.36, wherein the first preset time period is set according to specific requirements, which can be the same as or different from the preset time period, for example, the time point at which the content push device obtains the multiple user features of the training user can be determined as the end time point of the first preset time period, and the length of the first preset time period can be set according to requirements. For example, if the length of the first preset time period is 2 hours, and the time point at which the content push device obtains the multiple user features of the training user is 16:00, then the first preset time period is from 14:00 to 16:00. The push side feature of the training user is used to represent the click feature of the training user on the historical push content of each content type pushed to the training user within the first preset time period, which can include content type feature of the historical push content, click feature on the historical push content of the daily report type, click feature on the historical push content of the hot spot type, and click feature on the historical push content of the social type, etc.; wherein the content type feature of the historical push content indicates the content type of the historical push content pushed to the training user within the first preset time period, including the daily report type, the hot spot type, the social type, etc., the click feature on the historical push content of the daily report type represents the click rate of the training user on the historical push content of the daily report type pushed to the training user within the first preset time period, which can be represented by a floating point number value, for example, 0.32; the click feature on the historical push content of the hot spot type represents the click rate of the training user on the historical push content of the hot spot type pushed to the training user within the first preset time period, which can be represented by a floating point number value, for example, 0.22; and the click feature on the historical push content of the social type represents the click rate of the training user on the historical push content of the social type pushed to the training user within the first preset time period, which can be represented by a floating point number value, for example, 0.54.

[0083] The social side feature of the training user is used to represent the social feature of the training user, which can include a social friend number feature, a male proportion in social friends feature, a female proportion in social friends feature, an age feature of social friends, and a click feature of social friends, etc. The social friend number feature represents the number of social friends of the training user, which can be represented by an integer value, such as 20. The male proportion in social friends feature represents the proportion of males in the social friends of the training user, which can be represented by a floating point number value, such as 0.65. The female proportion in social friends feature represents the proportion of females in the social friends of the training user, which can be represented by a floating point number value, such as 0.35. The age feature of social friends represents the average age of the social friends of the training user, which can be represented by an integer value, such as 27. The click feature of social friends represents the click rate of the historical push content pushed to the social friends of the training user in the first preset time period, which can be represented by a floating point number value, such as 0.27. The context feature of the training user is used to represent the state feature of the training user, which can include a time feature, a device feature, a network feature, a power feature, a motion feature, etc. The time feature represents the time period in which the training user is located, i.e. the time period in which the training user browses the content detail page of the historical push content, which can be, for example, morning, afternoon, evening, early morning, etc. The device feature represents the device model of the terminal device when the training user browses the content detail page of the historical push content, such as device model 1, device model 2, etc. The network feature represents the network state of the terminal device when the training user browses the content detail page of the historical push content, which can be, for example, wireless network (WIFI), 2G network, 3G network, 4G network, 5G network, etc. The power feature represents the power state of the terminal device when the training user browses the content detail page of the historical push content, which can be, for example, full, sufficient, low, very low. The motion feature represents whether the training user is in a motion state when browsing the content detail page of the historical push content, which can be obtained by the corresponding motion function module in the target application. As shown in FIG. 8, Figure 10 FIG. 8 is a schematic diagram of a plurality of user features of a training user provided by an embodiment of the present application.

[0084] In S802, the user side feature of the training user and the push side feature of the training user are coupled and processed by a first coupler in the content type prediction model to obtain a first coupled representation vector of the training user.

[0085] In an embodiment, the content pushing device performs feature coupling processing on the user-side features of the training user and the push-side features of the training user through a first coupler in the content type prediction model to obtain a first coupled feature vector of the training user, so as to learn the interaction influence between the user-side features of the training user and the push-side features of the training user. The first coupler in the content type prediction model can be a coupler based on a factorization machine (FM), a coupler based on an enhanced sequential inference model (ESIM), a coupler based on a crossnet (CrossNet), or any coupler capable of performing feature coupling processing on the user-side features of the training user and the push-side features of the training user.

[0086] In an embodiment, if the first coupler in the content type prediction model is a coupler based on an FM, the content pushing device performs feature coupling processing on the user-side features of the training user and the push-side features of the training user through the first coupler in the content type prediction model to obtain a first coupled feature vector of the training user, which can include: performing feature interaction calculation on each two different feature values in the multiple feature values included in the user-side features of the training user and the push-side features of the training user through the first coupler in the content type prediction model to obtain a first feature interaction calculation result of the training user; performing linear summation calculation on each feature value in the multiple feature values included in the user-side features of the training user and the push-side features of the training user to obtain a first linear summation calculation result of the training user; and performing summation calculation on the first feature interaction calculation result of the training user and the first linear summation calculation result of the training user to obtain the first coupled feature vector of the training user.

[0087] Further, the content pushing device obtains the first coupling feature vector of the training user by performing feature coupling processing on the user-side feature of the training user and the pushing-side feature of the training user through the first coupler in the content type prediction model. That is, before the content pushing device obtains the first coupling feature vector of the training user by performing feature coupling processing on the user-side feature of the training user and the pushing-side feature of the training user through the first coupler in the content type prediction model, the content pushing device also performs feature embedding processing on the user-side feature of the training user and the pushing-side feature of the training user through the content type prediction model to obtain the feature representation vector of the user-side feature of the training user and the feature representation vector of the pushing-side feature of the training user. Further, the content pushing device obtains the first coupling feature vector of the training user by performing feature coupling processing on the user-side feature of the training user and the pushing-side feature of the training user through the first coupler in the content type prediction model, which can include: the content pushing device obtains the first coupling feature vector of the training user by performing feature coupling processing on the feature representation vector of the user-side feature of the training user and the feature representation vector of the pushing-side feature of the training user through the first coupler in the content type prediction model.

[0088] Further, the content pushing device obtains the first coupling feature vector of the training user by performing feature coupling processing on the user-side feature of the training user and the pushing-side feature of the training user through the first coupler in the content type prediction model. That is, before the content pushing device obtains the first coupling feature vector of the training user by performing feature coupling processing on the user-side feature of the training user and the pushing-side feature of the training user through the first coupler in the content type prediction model, the content pushing device also performs feature embedding processing on the user-side feature of the training user and the pushing-side feature of the training user through the content type prediction model to obtain the feature representation vector of the user-side feature of the training user and the feature representation vector of the pushing-side feature of the training user. Further, the content pushing device obtains the first coupling feature vector of the training user by performing feature coupling processing on the user-side feature of the training user and the pushing-side feature of the training user through the first coupler in the content type prediction model, which can include: the content pushing device obtains the first coupling feature vector of the training user by performing feature coupling processing on the feature representation vector of the user-side feature of the training user and the feature representation vector of the pushing-side feature of the training user through the first coupler in the content type prediction model.

[0089] In an embodiment, if the feature representation vector of the user-side feature of the training user and the feature representation vector of the push-side feature of the training user are spliced to obtain a first spliced feature representation vector of the training user, the first spliced feature representation vector of the training user is the spliced feature representation vector to be processed by the first coupler in the content type prediction model; then, the content push device can perform feature coupling processing on the first spliced feature vector of the training user through the first coupler in the content type prediction model to realize feature coupling processing on the feature representation vector of the user-side feature of the training user and the feature representation vector of the push-side feature of the training user, at this time, the first coupled representation vector of the training user can be determined by formula (2):

[0090]

[0091] Wherein, n is the number of feature values in the first spliced feature representation vector of the training user, i and j are integer variables; w0, w i and w ij are parameters of the first coupler in the content type prediction model; x i is the i-th feature value in the first spliced feature representation vector of the training user, x j is the j-th feature value in the first spliced feature representation vector of the training user, The value of is the first feature interaction calculation result of the training user, is the first linear summation calculation result of the training user.

[0092] S803, through the second coupler in the content type prediction model, the user-side feature of the training user and the social-side feature of the training user are coupled to obtain a second coupled representation vector of the training user.

[0093] S804, through the third coupler in the content type prediction model, the user-side feature of the training user and the context feature of the training user are coupled to obtain a third coupled representation vector of the training user.

[0094] Wherein, the implementation process of step S803 and step S804 is similar to the implementation process of step S802, which will not be repeated here, and step S802 to step S804 does not have a strict execution order, and can be executed in parallel.

[0095] S805, the first coupled representation vector of the training user, the second coupled representation vector of the training user and the third coupled representation vector of the training user are analyzed to obtain a type representation vector of the training user.

[0096] For example, Figure 11As shown, another schematic diagram of predicting the content type that the training user pays attention to through the content type prediction model provided by the embodiment of the present application is shown, the content type prediction model includes an input layer, an embedding layer, a network mapping layer, and an output layer, wherein the network mapping layer includes a first coupler, a second coupler, a third coupler, and a fully connected layer; the content pushing device receives the user side feature of the training user, the pushing side feature of the training user, the social side feature of the training user, and the context feature of the training user through the input layer, and inputs the user side feature of the training user, the pushing side feature of the training user, the social side feature of the training user, and the context feature of the training user to the embedding layer through the input layer; the user side feature of the training user, the pushing side feature of the training user, the social side feature of the training user, and the context feature of the training user are respectively subjected to feature embedding processing through the embedding layer, to obtain a plurality of feature representation vectors of the training user (the feature representation vector of the user side feature of the training user, the feature representation vector of the pushing side feature of the training user, the feature representation vector of the social side feature of the training user, and the feature representation vector of the context feature of the training user); the feature representation vector of the user side feature of the training user and the feature representation vector of the pushing side feature of the training user are subjected to feature coupling processing through the first coupler in the network mapping layer, to obtain a first coupling representation vector of the training user; the feature representation vector of the user side feature of the training user and the feature representation vector of the social side feature of the training user are subjected to feature coupling processing through the second coupler in the network mapping layer, to obtain a second coupling representation vector of the training user; the feature representation vector of the user side feature of the training user and the feature representation vector of the context feature of the training user are subjected to feature coupling processing through the third coupler in the network mapping layer, to obtain a third coupling representation vector of the training user; the first coupling representation vector of the training user, the second coupling representation vector of the training user, and the third coupling representation vector of the training user are subjected to feature analysis processing through the fully connected layer in the network mapping layer, to obtain a type representation vector of the training user; further, the content type that the training user pays attention to is subjected to prediction processing based on the type representation vector of the training user through the output layer, to predict the content type that the training user pays attention to, to obtain the predicted content type that the training user pays attention to, that is, step S806 is executed.

[0097] In one embodiment, after the content push device obtains the first coupled representation vector, the second coupled representation vector, and the third coupled representation vector of the training user through a content type prediction model, it can further perform feature analysis processing on the user-side features, the first coupled representation vector, the second coupled representation vector, and the third coupled representation vector of the training user to obtain the type representation vector of the training user. Specifically, this feature analysis is achieved by performing feature analysis processing on the user-side feature representation vector, the first coupled representation vector, the second coupled representation vector, and the third coupled representation vector of the training user. Figure 12 The diagram illustrates another method for predicting the content types a training user is interested in using a content type prediction model, as provided in this embodiment. A fully connected layer can be used to perform feature analysis on the user-side feature representation vector, the first coupled representation vector, the second coupled representation vector, and the third coupled representation vector of the training user to obtain the training user's type representation vector. The push-side features, social-side features, and contextual features of the training user are coupled with the user-side features using different couplers because the user-side features reflect the basic information of the training user. Similarly, the first coupled representation vector, the second coupled representation vector, and the third coupled representation vector of the training user are analyzed and processed with the user-side features because the user-side features reflect the basic information of the training user.

[0098] S806, based on the type representation vector of the training user, performs prediction processing on the content type that the training user is interested in, predicts the content type that the training user is interested in, and obtains the predicted content type that the training user is interested in.

[0099] S807, the content type prediction model is trained based on the predicted content types followed by the training users and the labeled content types followed by the training users, to obtain the trained content prediction model.

[0100] In the embodiments of the present application, the content pushing device trains the content type prediction model based on the user side features of the training user, the pushing side features of the training user, the social side features of the training user, the context features of the training user, and the annotated content type that the training user is interested in, so that the trained content type prediction model can accurately predict the content type that the target user is interested in based on the user side features of the target user, the pushing side features of the target user, the social side features of the target user, and the context features of the target user, that is, the content type that different users are interested in can be personalized predicted based on multiple user features of different users.

[0101] Based on the above system embodiments and method embodiments, the embodiments of the present application provide a content pushing method, as shown in Figure 13 The content pushing method shown in FIG. 13 can be executed by the content pushing device. Figure 13 The content pushing method shown in FIG. 13 can be executed by the content pushing device. Figure 13 The content pushing method shown in FIG. 13 can include the following steps:

[0102] S1301, obtaining multiple user features of a target user.

[0103] The target user is any user in the target application. The multiple user features of the target user are user features of the target user in different dimensions. The multiple user features of the target user are used to process through the trained content type prediction model to predict the target content type that the target user is interested in. The predicted target content type that the target user is interested in is one of the preset content types. The multiple user features of the target user correspond to the multiple user features required by the trained content type prediction model. For example, if the trained content type prediction model is obtained by training the content type prediction model based on the user side features of the training user, the pushing side features of the training user, the social side features of the training user, and the context features of the training user, then the multiple user features of the target user are the user side features of the target user, the pushing side features of the target user, the social side features of the target user, and the context features of the target user.

[0104] In an embodiment, if the plurality of user features of the target user are the user-side feature of the target user, the push-side feature of the target user, the social-side feature of the target user, and the context feature of the target user, the click feature in the user-side feature of the target user represents the click rate of the target user on the historical push content pushed to the target user in the second preset time period, which can be represented by a floating-point number value, for example, 0.26, wherein, the second preset time period is set according to specific requirements, for example, the time point at which the content push device obtains the plurality of user features of the target user can be determined as the end time point of the second preset time period, the length of the second preset time period can be set according to requirements, for example, if the length of the second preset time period is 0.5 hours, and the time point at which the content push device obtains the plurality of user features of the target user is 16:00, then the second preset time period is from 15:30 to 16:00; the click feature on the historical push content of the daily report type in the push-side feature of the target user represents the click rate of the target user on the historical push content of the daily report type pushed to the target user in the second preset time period, which can be represented by a floating-point number value, for example, 0.32; the click feature on the historical push content of the hot spot type represents the click rate of the target user on the historical push content of the hot spot type pushed to the target user in the second preset time period, which can be represented by a floating-point number value, for example, 0.22; the click feature on the historical push content of the social type represents the click rate of the target user on the historical push content of the social type pushed to the target user in the second preset time period, which can be represented by a floating-point number value, for example, 0.54; the click feature of the social friend in the social-side feature of the target user represents the click rate of the social friend of the target user on the historical push content pushed to the social friend in the second preset time period, which can be represented by a floating-point number value, for example, 0.27.

[0105] In S1302, the plurality of user features of the target user are processed by the trained content type prediction model to obtain a type representation vector of the target user, and the content type concerned by the target user is predicted based on the type representation vector of the target user to obtain the target content type concerned by the target user.

[0106] In an embodiment, the content pushing device can perform feature fusion processing on the plurality of user features of the target user by using the trained content type prediction model to obtain a type representation vector of the target user, which can include: performing feature coupling processing on the user-side features of the target user and the push-side features of the target user by using a first coupler in the trained content type prediction model to obtain a first coupled representation vector of the target user; performing feature coupling processing on the user-side features of the target user and the social-side features of the target user by using a second coupler in the trained content type prediction model to obtain a second coupled representation vector of the target user; performing feature coupling processing on the user-side features of the target user and the context features of the target user by using a third coupler in the trained content type prediction model to obtain a third coupled representation vector of the target user; and performing feature analysis processing on the first coupled representation vector of the target user, the second coupled representation vector of the target user, and the third coupled representation vector of the target user to obtain the type representation vector of the target user.

[0107] Further, the content pushing device can perform feature coupling processing on the user-side features of the target user and the push-side features of the target user by using the first coupler in the trained content type prediction model to obtain the first coupled representation vector of the target user, which can include: performing feature interaction calculation on each two different feature values in the plurality of feature values included in the user-side features of the target user and the push-side features of the target user by using the first coupler in the trained content type prediction model to obtain a first feature interaction calculation result of the target user; performing linear summation calculation on each feature value in the plurality of feature values included in the user-side features of the target user and the push-side features of the target user to obtain a first linear summation calculation result of the target user; and performing summation calculation on the first feature interaction calculation result of the target user and the first linear summation calculation result of the target user to obtain the first coupled representation vector of the target user. The content pushing device obtains the second coupled representation vector of the target user and the third coupled representation vector of the target user in a similar manner as obtaining the first coupled representation vector of the target user, and thus details are not repeated here.

[0108] In a specific implementation, the content pushing device performs feature coupling processing on the user-side feature of the target user and the pushing-side feature of the target user through a first coupler in the trained content type prediction model to obtain a first coupled representation vector of the target user, which is obtained by processing a feature representation vector of the user-side feature of the target user and a feature representation vector of the pushing-side feature of the target user; performs feature coupling processing on the user-side feature of the target user and the social-side feature of the target user through a second coupler in the trained content type prediction model to obtain a second coupled representation vector of the target user, which is obtained by processing the feature representation vector of the user-side feature of the target user and a feature representation vector of the social-side feature of the target user; and performs feature coupling processing on the user-side feature of the target user and the context feature of the target user through a third coupler in the trained content type prediction model to obtain a third coupled representation vector of the target user, which is obtained by processing the feature representation vector of the user-side feature of the target user and a feature representation vector of the context feature of the target user.

[0109] As Figure 14As shown, a schematic diagram of obtaining a target content type concerned by a target user through a trained content type prediction model is provided, the trained content type prediction model comprises an input layer, an embedding layer, a network mapping layer and an output layer, wherein the network mapping layer comprises a first coupler, a second coupler, a third coupler and a full connection layer; the content pushing device receives user side features of the target user, pushing side features of the target user, social side features of the target user and context features of the target user through the input layer, and inputs the user side features of the target user, the pushing side features of the target user, the social side features of the target user and the context features of the target user to the embedding layer through the input layer; the user side features of the target user, the pushing side features of the target user, the social side features of the target user and the context features of the target user are respectively processed by the embedding layer for feature embedding, to obtain a plurality of feature representation vectors of the target user (respectively, a feature representation vector of the user side features of the target user, a feature representation vector of the pushing side features of the target user, a feature representation vector of the social side features of the target user and a feature representation vector of the context features of the target user); the feature representation vector of the user side features of the target user and the feature representation vector of the pushing side features of the target user are processed by the first coupler in the network mapping layer for feature coupling, to obtain a first coupled representation vector of the target user; the feature representation vector of the user side features of the target user and the feature representation vector of the social side features of the target user are processed by the second coupler in the network mapping layer for feature coupling, to obtain a second coupled representation vector of the target user; the feature representation vector of the user side features of the target user and the feature representation vector of the context features of the target user are processed by the third coupler in the network mapping layer for feature coupling, to obtain a third coupled representation vector of the target user; the first coupled representation vector of the target user, the second coupled representation vector of the target user and the third coupled representation vector of the target user are processed by the full connection layer in the network mapping layer for feature analysis, to obtain a type representation vector of the target user; the content type concerned by the target user is processed by the output layer based on the type representation vector of the target user, to obtain a target content type concerned by the target user.

[0110] The content pushing device predicts the content type that the target user pays attention to based on the type representation vector of the target user through the output layer. When the target content type that the target user pays attention to is predicted, the type representation vector of the target user is mapped to the probability value of the preset content type through the output layer. When the maximum probability value is obtained, the preset content type corresponding to the maximum probability value is determined as the target content type that the target user pays attention to. For example, if the preset content type is a daily report type, a hot spot type and a social type, the type representation vector of the target user is mapped to the probability values of the daily report type, the hot spot type and the social type, which are 0.5%, 99% and 0.5% respectively. The target content type that the target user pays attention to is obtained as the hot spot type.

[0111] In one embodiment, after the content pushing device obtains the first coupling representation vector of the training user, the second coupling representation vector of the target user and the third coupling representation vector of the target user through the content type prediction model, the user side feature of the target user, the first coupling representation vector of the target user, the second coupling representation vector of the target user and the third coupling representation vector of the target user can be analyzed and processed to obtain the type representation vector of the target user. In a specific implementation, the user side feature of the target user, the first coupling representation vector of the target user, the second coupling representation vector of the target user and the third coupling representation vector of the target user can be analyzed and processed by analyzing and processing the feature representation vector of the user side feature of the target user, the first coupling representation vector of the target user, the second coupling representation vector of the target user and the third coupling representation vector of the target user. Figure 15 As shown in FIG. 13B, another schematic diagram for obtaining the target content type that the target user pays attention to through the trained content type prediction model is provided in the embodiment of the present application. The feature representation vector of the user side feature of the target user, the first coupling representation vector of the target user, the second coupling representation vector of the target user and the third coupling representation vector of the target user can be analyzed and processed through the fully connected layer to obtain the type representation vector of the target user.

[0112] S1303, determining the target to-be-pushed content from the plurality of to-be-pushed contents.

[0113] The to-be-pushed content is related to the service provided by the target application. For example, if the target application is a news application, the to-be-pushed content is news content. The target to-be-pushed content matches the target content type that the target user pays attention to, and the target to-be-pushed content is used to push to the target user.

[0114] In an embodiment, the content pushing device determines the target to-be-pushed content from the plurality of to-be-pushed contents, which can include: determining a first candidate pushed content matching the target content type that the target user is interested in from the plurality of to-be-pushed contents, and selecting the target to-be-pushed content matching the target user from the first candidate pushed content. Further, the content pushing device selects the target to-be-pushed content matching the target user from the first candidate pushed content, which can include: determining a target historical pushed content under the target content type, the target historical pushed content under the target content type referring to a historical pushed content under the target content type pushed to the target user in a third preset time period, and the historical pushed content being browsed by the target user; determining a target content label of the target historical pushed content; and selecting the target to-be-pushed content matching the target content label of the target historical pushed content from the first candidate pushed content according to the target content label of the target historical pushed content and a content label of the first candidate pushed content.

[0115] The third preset time period is set according to specific requirements, for example, a time point corresponding to a plurality of user features of the target user acquired by the content pushing device can be determined as an end time point of the third preset time period, and the length of the third preset time period can be set according to requirements. For example, if the length of the third preset time period is 1 hour, and the time point corresponding to the plurality of user features of the target user acquired by the content pushing device is 15:00, then the historical pushed content under the target content type pushed to the target user in the third preset time period is the historical pushed content under the target content type pushed to the target user from 14:00 to 15:00.

[0116] In an embodiment, the content label is a label describing the content theme of the content, such as an entertainment, a movie, a gossip, a cute pet, and the like. The content pushing device determining the target content label of the target historical pushing content can include: obtaining the content labels of the target historical pushing content, and determining the target content label of the target historical pushing content from the obtained content labels of the target historical pushing content. The target content label of the target historical pushing content can be determined from the obtained content labels of the target historical pushing content according to different requirements. For example, the target content label of the target historical pushing content can be determined as the content labels with top quantities in the content labels of the target historical pushing content. For example, if the target historical pushing content is historical pushing content 1, historical pushing content 2, and historical pushing content 3, the content label of the historical pushing content 1 is content label 1, content label 2, and content label 3, the content label of the historical pushing content 2 is content label 1, content label 2, content label 4, and content label 5, and the content label of the historical pushing content 3 is content label 1, content label 2, content label 3, and content label 6, if the target content label of the target historical pushing content is determined as the content labels with top 3 quantities in the content labels of the target historical pushing content, the determined target content label of the target historical pushing content is content label 1, content label 2, and content label 3.

[0117] In an embodiment, the content pushing device selecting the target pushing content matching the target content label of the target historical pushing content from the first candidate pushing content according to the target content label of the target historical pushing content and the content labels of the first candidate pushing content can include: calculating the label similarity between the target content label of the target historical pushing content and the content labels of the first candidate pushing content, and determining the first candidate pushing content corresponding to the maximum label similarity as the target pushing content. Optionally, the label similarity between the target content label of the target historical pushing content and the content labels of the first candidate pushing content can be a cosine similarity.

[0118] In an embodiment, the content pushing device determines the target to-be-pushed content from the plurality of to-be-pushed contents, which can include: selecting a second candidate pushed content matching the target user from the plurality of to-be-pushed contents, and performing format conversion processing on the second candidate pushed content to obtain the target to-be-pushed content, the format of the target to-be-pushed content matching the target content type that the target user pays attention to. The process of selecting the second candidate pushed content matching the target user from the plurality of to-be-pushed contents is similar to that of selecting the target to-be-pushed content matching the target user from the first candidate pushed content, and will not be described here. Further, the content pushing device performs format conversion processing on the second candidate pushed content to obtain the target to-be-pushed content, in order to convert the format of the second candidate pushed content to a format matching the target content type that the target user pays attention to, i.e., to convert the format of the second candidate pushed content to a format that the target user is interested in, which can improve the attractiveness of the second candidate pushed content to the target user. For example, if the preset content type corresponding to the trained content type prediction model is a text type, a video type, and an audio type, if the content pushing device determines that the second candidate pushed content is a text type content, but the target content type that the target user pays attention to is an audio type, then the text type content can be converted to an audio format expression through a speech synthesis technology (also known as a text-to-speech (TTS) technology), because the audio format content is the target content type that the target user pays attention to, so the attractiveness of the second candidate pushed content to the target user can be improved, and the click rate, user experience, user activity, and user stickiness of the user can be further improved. For another example, if the content pushing device determines that the second candidate pushed content is an audio type content, but the target content type that the target user pays attention to is a text type, then the audio type content can be converted to a text format expression through a speech recognition technology.

[0119] In an embodiment, after the content pushing device obtains the target to-be-pushed content, the content pushing device can push the target to-be-pushed content to the terminal device of the target user, for example, the content pushing device can push the target to-be-pushed content to a target application running in the terminal device of the target user, so that the target to-be-pushed content can be displayed based on the target application; or the content pushing device can push the target to-be-pushed content to the terminal device of the target user in the form of a notification information, so that the target to-be-pushed content can be displayed based on the pop-up window of the notification information.

[0120] In the embodiment of the present application, the content pushing device can obtain a plurality of user features of a target user, the plurality of user features of the target user being user features of the target user in different dimensions; the plurality of user features of the target user are processed by feature fusion through the trained content type prediction model to obtain a type representation vector of the target user, and the content type concerned by the target user is predicted based on the type representation vector of the target user to obtain a target content type concerned by the target user; and the target content to be pushed is determined from a plurality of content to be pushed, the target content to be pushed matching the target content type concerned by the target user, and the target content to be pushed being used to push to the target user. The content pushing device can predict the target content type concerned by the user according to the plurality of user features of the user, and then determine the target content to be pushed matching the target content type concerned by the user from the content to be pushed, so as to realize personalized prediction of the content type concerned by different users, and then determine the target content to be pushed from the content to be pushed according to the target content type concerned by the user predicted, so as to realize personalized determination of the content to be pushed, and then realize personalized pushing of the content, so as to improve user experience, user activity and user stickiness.

[0121] Based on the above system embodiment and method embodiment, the embodiment of the present application provides a content pushing device. Referring to Figure 16 A structural schematic diagram of a content pushing device provided by the embodiment of the present application can include an acquisition unit 1601 and a processing unit 1602. Figure 16 The content pushing device shown can run the following units:

[0122] The acquisition unit 1601 is configured to obtain a plurality of user features of a target user, the plurality of user features of the target user being user features of the target user in different dimensions;

[0123] The processing unit 1602 is configured to process the plurality of user features of the target user by feature fusion through a trained content type prediction model to obtain a type representation vector of the target user, and predict a content type concerned by the target user based on the type representation vector of the target user to obtain a target content type concerned by the target user.

[0124] The processing unit 1602 is further configured to determine a target content to be pushed from a plurality of content to be pushed, the target content to be pushed matching the target content type concerned by the target user, and the target content to be pushed being used to push to the target user.

[0125] In one embodiment, when the processing unit 1602 determines the target content to be pushed from the plurality of content to be pushed, the following steps are specifically performed: In one embodiment, when the processing unit 1602 determines the target content to be pushed from the plurality of content to be pushed, the following steps are specifically performed:

[0126] determining, from the plurality of to-be-pushed contents, a first candidate pushed content matching a target content type that the target user pays attention to, and selecting, from the first candidate pushed content, a target to-be-pushed content matching the target user.

[0127] In an embodiment, when determining the target to-be-pushed content from the plurality of to-be-pushed contents, the processing unit 1602 specifically performs the following operations:

[0128] selecting, from the plurality of to-be-pushed contents, a second candidate pushed content matching the target user, and performing format conversion processing on the second candidate pushed content to obtain the target to-be-pushed content, wherein the format of the target to-be-pushed content matches the target content type that the target user pays attention to.

[0129] In an embodiment, the plurality of user features of the target user include: a user-side feature of the target user, a push-side feature of the target user, a social-side feature of the target user, and a context feature of the target user.

[0130] When the processing unit 1602 performs feature fusion processing on the plurality of user features of the target user by using the trained content type prediction model to obtain a type representation vector of the target user, the processing unit 1602 specifically performs the following operations:

[0131] performing, by a first coupler in the trained content type prediction model, feature coupling processing on the user-side feature of the target user and the push-side feature of the target user to obtain a first coupled representation vector of the target user.

[0132] performing, by a second coupler in the trained content type prediction model, feature coupling processing on the user-side feature of the target user and the social-side feature of the target user to obtain a second coupled representation vector of the target user.

[0133] performing, by a third coupler in the trained content type prediction model, feature coupling processing on the user-side feature of the target user and the context feature of the target user to obtain a third coupled representation vector of the target user.

[0134] performing feature analysis processing on the first coupled representation vector of the target user, the second coupled representation vector of the target user, and the third coupled representation vector of the target user to obtain the type representation vector of the target user.

[0135] In an embodiment, when the processing unit 1602 performs feature analysis processing on the first coupling feature vector of the target user, the second coupling feature vector of the target user, and the third coupling feature vector of the target user to obtain the type feature vector of the target user, the following operations are specifically performed:

[0136] The user-side features of the target user, the first coupling feature vector of the target user, the second coupling feature vector of the target user, and the third coupling feature vector of the target user are subjected to feature analysis processing to obtain the type feature vector of the target user.

[0137] In an embodiment, when the processing unit 1602 performs feature coupling processing on the user-side features of the target user and the push-side features of the target user through the first coupler in the trained content type prediction model to obtain the first coupling feature vector of the target user, the following operations are specifically performed:

[0138] The feature interaction calculation is performed on each two different feature values in the plurality of feature values included in the user-side features of the target user and the push-side features of the target user through the first coupler in the trained content type prediction model to obtain the first feature interaction calculation result of the target user.

[0139] The linear summation calculation is performed on each feature value in the plurality of feature values included in the user-side features of the target user and the push-side features of the target user to obtain the first linear summation calculation result of the target user.

[0140] The summation calculation is performed on the first feature interaction calculation result of the target user and the first linear summation calculation result of the target user to obtain the first coupling feature vector of the target user.

[0141] In an embodiment, the acquisition unit 1601 is further configured to acquire a training sample, wherein the training sample includes a plurality of user features of a training user and a labeled content type that the training user pays attention to, and the plurality of user features of the training user are user features of the training user in different dimensions.

[0142] The processing unit 1602 is further configured to take the plurality of user features of the training user as model input, take the labeled content type that the training user pays attention to as expected output, train a content type prediction model based on the training sample, and obtain the trained content type prediction model.

[0143] In an embodiment, when the acquisition unit 1601 acquires a training sample, the following steps are specifically performed:

[0144] determine historical push content pushed to a user within a preset time period, and the user who browses a content detail page of the historical push content;

[0145] determine historical push content pushed to a user within a preset time period, and the user who browses a content detail page of the historical push content;

[0146] determine historical push content pushed to a user within a preset time period, and the user who browses a content detail page of the historical push content;

[0147] determine historical push content pushed to a user within a preset time period, and the user who browses a content detail page of the historical push content;

[0148] According to one embodiment of the present application, Figure 2 、 Figure 6 、 Figure 8 and Figure 13 The steps involved in the content pushing method shown in Figure 16 The steps involved in the content pushing method shown in Figure 2 The steps involved in the content pushing method shown in Figure 16 The steps involved in the content pushing method shown in Figure 2 The steps involved in the content pushing method shown in Figure 16 The steps involved in the content pushing method shown in Figure 6 The steps involved in the content pushing method shown in Figure 16 The steps involved in the content pushing method shown in Figure 6 The steps involved in the content pushing method shown in Figure 16 The steps involved in the content pushing method shown in Figure 8 The steps involved in the content pushing method shown in Figure 16 The steps involved in the content pushing method shown in Figure 8 The steps involved in the content pushing method shown in Figure 16 The steps involved in the content pushing method shown in Figure 13 The steps involved in the content pushing method shown in Figure 16 The steps involved in the content pushing method shown in Figure 13 The steps involved in the content pushing method shown in Figure 16 The steps involved in the content pushing method shown in

[0149] According to another embodiment of the present application, Figure 16The units in the content pushing apparatus shown can be combined into one or several other units respectively or all, or some of the units can be further split into a plurality of units with smaller functions to constitute, which can achieve the same operation without affecting the implementation of the technical effects of the embodiments of the present application. The above units are divided based on logical functions, and in actual application, the functions of one unit can also be implemented by a plurality of units, or the functions of a plurality of units are implemented by one unit. In other embodiments of the present application, the content pushing apparatus divided based on logical functions can also include other units, and in actual application, these functions can also be assisted by other units to be implemented, and can be implemented by a plurality of units in cooperation.

[0150] According to another embodiment of the present application, the content pushing apparatus as shown in Figure 2 、 Figure 6 、 Figure 8 and Figure 13 the corresponding method shown in can be constructed, and the content pushing method of the embodiments of the present application can be implemented. The computer program can be recorded on, for example, a computer readable storage medium, and loaded into the above-mentioned computing device through the computer readable storage medium, and run therein. Figure 16

[0151] In the embodiments of the present application, the acquisition unit 1601 can acquire a plurality of user features of a target user, the plurality of user features of the target user being user features of the target user in different dimensions; the processing unit 1602 can perform feature fusion processing on the plurality of user features of the target user through the trained content type prediction model, obtain a type representation vector of the target user, and perform prediction processing on the content type that the target user pays attention to based on the type representation vector of the target user, and predict the target content type that the target user pays attention to; determine a target to-be-pushed content in a plurality of to-be-pushed contents, the target to-be-pushed content matching the target content type that the target user pays attention to, and the target to-be-pushed content being used to push to the target user. The target content type that the user pays attention to can be predicted according to the plurality of user features of the user, and then the target to-be-pushed content matching the target content type that the user pays attention to is determined in the to-be-pushed content, that is, the personalized prediction of the content type that the user pays attention to can be realized for different users, and then the target to-be-pushed content is determined from the to-be-pushed content according to the target content type that the user pays attention to predicted, which can realize the personalized determination of the to-be-pushed content, and then the personalized pushing of the content can be realized to improve the user experience, user activity and user stickiness.

[0152] ​Based on the above method embodiments and device embodiments, the application further provides a content pushing device. Referring to Figure 17 FIG. 1 is a structural schematic diagram of a content pushing device according to an embodiment of the application. Figure 17 The content pushing device shown in FIG. 1 can at least include a processor 1701, an input interface 1702, an output interface 1703, and a computer storage medium 1704. The processor 1701, the input interface 1702, the output interface 1703, and the computer storage medium 1704 can be connected through a bus or other means.

[0153] The computer storage medium 1704 can be stored in the memory of the content pushing device, and the computer storage medium 1704 is used to store a computer program, the computer program including program instructions, and the processor 1701 is used to execute the program instructions stored in the computer storage medium 1704. The processor 1701 (or CPU (Central Processing Unit, central processor)) is the computing core and control core of the content pushing device, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions to implement the content pushing method flow or the corresponding function described above.

[0154] The embodiment of the application further provides a computer storage medium (Memory), which is a memory device in the content pushing device and is used to store programs and data. It can be understood that the computer storage medium herein can include the built-in storage medium in the terminal, and of course can also include the expansion storage medium supported by the terminal. The computer storage medium provides a storage space, and the storage space stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor 1701 are also stored in the storage space, and the instructions can be one or more computer programs (including program codes). It should be noted that the computer storage medium herein can be a high-speed random access memory (RAM) memory, or a non-volatile memory such as at least one disk memory; and optionally can be at least one computer storage medium located away from the aforementioned processor.

[0155] In one embodiment, one or more instructions stored in the computer storage medium can be loaded and executed by the processor 1701 and the input interface 1702 to implement the above-mentioned related Figure 2 、 Figure 6 、 Figure 8 and Figure 13corresponding steps of the method in the content pushing method embodiment, in a specific implementation, one or more instructions in the computer storage medium are loaded and executed by the processor 1701 and the input interface 1702 as follows:

[0156] The input interface 1702 is configured to obtain a plurality of user features of a target user, the plurality of user features of the target user being user features of the target user in different dimensions.

[0157] The processor 1701 is configured to perform feature fusion processing on the plurality of user features of the target user by using the trained content type prediction model, to obtain a type representation vector of the target user, and perform prediction processing on a content type that the target user pays attention to based on the type representation vector of the target user, to obtain a target content type that the target user pays attention to.

[0158] The processor 1701 is further configured to determine a target to-be-pushed content from a plurality of to-be-pushed contents, the target to-be-pushed content matching the target content type that the target user pays attention to, and the target to-be-pushed content being used for pushing to the target user.

[0159] In one embodiment, when the processor 1701 determines the target to-be-pushed content from the plurality of to-be-pushed contents, the processor 1701 specifically performs the following operations:

[0160] A first candidate to-be-pushed content matching the target content type that the target user pays attention to is determined from the plurality of to-be-pushed contents, and a target to-be-pushed content matching the target user is selected from the first candidate to-be-pushed content.

[0161] In one embodiment, when the processor 1701 determines the target to-be-pushed content from the plurality of to-be-pushed contents, the processor 1701 specifically performs the following operations:

[0162] A second candidate to-be-pushed content matching the target user is selected from the plurality of to-be-pushed contents, the second candidate to-be-pushed content is processed by format conversion to obtain the target to-be-pushed content, and the format of the target to-be-pushed content matches the target content type that the target user pays attention to.

[0163] In one embodiment, the plurality of user features of the target user include user-side features of the target user, push-side features of the target user, social-side features of the target user, and context features of the target user.

[0164] When the processor 1701 performs feature fusion processing on the plurality of user features of the target user by using the trained content type prediction model to obtain the type representation vector of the target user, the processor 1701 specifically performs the following operations:

[0165] The user-side feature of the target user and the push-side feature of the target user are coupled by a first coupler in the trained content type prediction model to obtain a first coupled representation vector of the target user;

[0166] The user-side feature of the target user and the social-side feature of the target user are coupled by a second coupler in the trained content type prediction model to obtain a second coupled representation vector of the target user;

[0167] The user-side feature of the target user and the context feature of the target user are coupled by a third coupler in the trained content type prediction model to obtain a third coupled representation vector of the target user;

[0168] The first coupled representation vector of the target user, the second coupled representation vector of the target user, and the third coupled representation vector of the target user are analyzed to obtain a type representation vector of the target user.

[0169] In one implementation, when the processor 1701 analyzes the first coupled representation vector of the target user, the second coupled representation vector of the target user, and the third coupled representation vector of the target user to obtain a type representation vector of the target user, the following operations are specifically performed:

[0170] The user-side feature of the target user, the first coupled representation vector of the target user, the second coupled representation vector of the target user, and the third coupled representation vector of the target user are analyzed to obtain a type representation vector of the target user.

[0171] In one embodiment, when the processor 1701 couples the user-side feature of the target user and the push-side feature of the target user by the first coupler in the trained content type prediction model to obtain the first coupled representation vector of the target user, the following operations are specifically performed:

[0172] Each of the two different feature values included in the user-side feature of the target user and the push-side feature of the target user is calculated by the first coupler in the trained content type prediction model to obtain a first feature interaction calculation result of the target user;

[0173] Each of the feature values included in the user-side feature of the target user and the push-side feature of the target user is linearly summed to obtain a first linear summation calculation result of the target user;

[0174] The first feature interaction calculation result of the target user is summed with the first linear summation calculation result of the target user to obtain a first coupling representation vector of the target user.

[0175] In an embodiment, the input interface 1702 is further configured to obtain a training sample, the training sample including a plurality of user features of a training user and a labeled content type that the training user pays attention to, the plurality of user features of the training user being user features of the training user in different dimensions.

[0176] The processor 1701 is further configured to take the plurality of user features of the training user as a model input, take the labeled content type that the training user pays attention to as an expected output, train a content type prediction model based on the training sample, and obtain the trained content type prediction model.

[0177] In an embodiment, when the input interface 1702 obtains the training sample, the following operations are specifically performed:

[0178] determine historical push content pushed to a user in a preset time period and the user who browses a content detail page of the historical push content;

[0179] take the determined user as a training user and obtain a plurality of user features of the training user;

[0180] label a content type of the historical push content as the content type that the training user pays attention to;

[0181] construct the training sample based on the plurality of user features of the training user and the labeled content type that the training user pays attention to.

[0182] Embodiments of the present application provide a computer program product or a computer program, which includes computer instructions stored in a computer readable storage medium. A processor of a content push device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to enable the content push device to perform the method embodiments as shown in Figure 2 、 Figure 6 、 Figure 8 or Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure 8 Figure 13 Figure 2 Figure 6 Figure . The computer readable storage medium can be a disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), or the like.

[0183] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A content push method, characterized in that, include: Acquire multiple user features of the target user, wherein the multiple user features of the target user are user features of the target user under different dimensions; The target user's multiple user characteristics include: user-side characteristics, push-side characteristics, social-side characteristics, and contextual characteristics. The social-side characteristics represent the target user's social features, including the number of social friends, the gender ratio of social friends, the age of social friends, and the click characteristics of social friends. The contextual characteristics represent the target user's state characteristics, including time characteristics, device characteristics, network characteristics, battery level characteristics, and motion characteristics. The multiple user features are input to the embedding layer through the input layer of the trained content type prediction model; the content type prediction model includes an input layer, an embedding layer, a network mapping layer and an output layer, and the network mapping layer includes a first coupler, a second coupler, a third coupler and a fully connected layer; if the preset content types corresponding to the content type prediction model are different, the user features used to train the content type prediction model will be different. The embedding layer performs feature embedding processing on the multiple user features respectively to obtain multiple feature representation vectors of the target user. The multiple feature representation vectors include feature representation vectors of user-side features, feature representation vectors of push-side features, feature representation vectors of social-side features, and feature representation vectors of context features. The feature representation vector of the user-side feature and the feature representation vector of the push-side feature are coupled through the first coupler in the network mapping layer to obtain the first coupled representation vector of the target user. The second coupler in the network mapping layer performs feature coupling processing on the feature representation vector of the user-side feature and the feature representation vector of the social-side feature to obtain the second coupled representation vector of the target user. The third coupler in the network mapping layer performs feature coupling processing on the feature representation vector of the user-side feature and the feature representation vector of the context feature to obtain the third coupled representation vector of the target user. The first coupled representation vector, the second coupled representation vector, and the third coupled representation vector are subjected to feature analysis processing through the fully connected layer in the network mapping layer to obtain the type representation vector of the target user. The output layer performs prediction processing on the type representation vector of the target user to predict the type of content that the target user is interested in, thereby predicting the type of target content that the target user is interested in. A target content to be pushed is determined from multiple content to be pushed, and the target content to be pushed is matched with the type of content that the target user is interested in. The target content to be pushed is used to push to the target user.

2. The method as described in claim 1, characterized in that, The step of determining the target content to be pushed from multiple content items includes: Among the plurality of content to be pushed, a first candidate content that matches the target content type that the target user is interested in is determined, and the target content to be pushed that matches the target user is selected from the first candidate content.

3. The method as described in claim 1, characterized in that, The step of determining the target content to be pushed from multiple content items includes: Select a second candidate push content that matches the target user from the plurality of push content, and perform format conversion processing on the second candidate push content to obtain the target push content. The format of the target push content matches the target content type that the target user is interested in.

4. The method as described in claim 1, characterized in that, The feature analysis processing of the first coupled representation vector, the second coupled representation vector, and the third coupled representation vector of the target user to obtain the type representation vector of the target user includes: Feature analysis processing is performed on the user-side features of the target user, the first coupled representation vector of the target user, the second coupled representation vector of the target user, and the third coupled representation vector of the target user to obtain the type representation vector of the target user.

5. The method as described in claim 1, characterized in that, The first coupler in the trained content type prediction model performs feature coupling processing on the user-side features and the push-side features of the target user to obtain the first coupled representation vector of the target user, including: Through the first coupler in the trained content type prediction model, feature interaction calculation is performed on every two different feature values ​​among the multiple feature values ​​included in the user-side features and push-side features of the target user to obtain the first feature interaction calculation result of the target user. A linear summation is performed on each of the multiple feature values ​​included in the user-side features and push-side features of the target user to obtain the first linear summation result of the target user. The first feature interaction calculation result of the target user is summed with the first linear summation calculation result of the target user to obtain the first coupled representation vector of the target user.

6. The method as described in claim 1, characterized in that, The method further includes: Obtain training samples, which include multiple user features of the training user and the content types that the training user is interested in, wherein the multiple user features of the training user are user features of the training user in different dimensions. The training user's multiple user features are used as model input, and the labeled content types that the training user is interested in are used as expected output. The content type prediction model is trained based on the training samples to obtain the trained content type prediction model.

7. The method as described in claim 6, characterized in that, The acquisition of training samples includes: Determine the historical push content pushed to users within a preset time period, and the users who browsed the content details page of the historical push content; The identified users are used as training users, and multiple user characteristics of the training users are obtained. The content type of the historical push content is labeled as the content type that the training user is interested in; The training samples are constructed based on multiple user features of the training users and the types of content that the training users are interested in, as labeled.

8. A content push device, characterized in that, include: The content push device includes an acquisition unit and a processing unit, wherein the acquisition unit and the processing unit are used to execute the content push method as described in any one of claims 1-7.

9. A computer storage medium, characterized in that, The computer storage medium stores computer program instructions, which, when executed by a processor, are used to perform the content push method as described in any one of claims 1-7.

10. A content push device, characterized in that, The content push device includes an input interface and an output interface, and also includes: A processor, adapted to implement one or more instructions; and, A computer storage medium storing one or more instructions, said one or more instructions being adapted to be loaded by the processor and executed as described in any one of claims 1-7.

11. A computer program product, characterized in that, The computer program product includes computer instructions, which, when executed by a processor, are used to perform the content push method as described in any one of claims 1-7.

Images