Information processing method and device, storage medium and computer equipment

By integrating personalized sub-networks to generate task weights based on user-specific characteristics, the model addresses the issue of poor prediction accuracy in existing systems, achieving improved accuracy and stability in user preference prediction.

CN119066252APending Publication Date: 2024-12-03TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310634358.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-31
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing recommendation systems fail to accurately predict user preferences due to the lack of consideration for personalized differences among users, leading to poor overall prediction performance.

Method used

A target prediction model is developed that incorporates personalized sub-networks to generate task weights based on user-specific characteristics, enhancing the model's ability to perform personalized learning for each user, thereby improving prediction accuracy and stability across multiple tasks.

Benefits of technology

The model effectively captures personalized user information, resulting in enhanced prediction accuracy and stability for different tasks, thus improving the overall performance of recommendation systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066252A_ABST
    Figure CN119066252A_ABST
Patent Text Reader

Abstract

The invention discloses an information processing method. The method comprises the steps of obtaining object information, content information and background information; inputting the object information, the content information and the background information into a target prediction model to predict a recommendation score of the target content, and outputting predicted scores of the target content under different tasks; the target prediction model is obtained by performing multi-task learning on a preset prediction network based on a sample target weight obtained by performing weight fusion on sample personalized weights and sample task weights corresponding to different tasks; the sample personalized weight corresponding to each task is generated by the personalized sub-network corresponding to each task based on the personalized characteristics of the object group. According to the method, the task weight generated by task gating is corrected through the personalized weight generated by each personalized sub-network, the multi-target modeling capability of the model in the personalized dimension is enhanced, and each task performs personalized learning on different object groups to enhance the prediction accuracy, so that the overall prediction performance of the model is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to an information processing method, apparatus, storage medium and computer equipment. Background Art

[0002] Recommendation systems are used to maximize user engagement and satisfaction in online applications by combining various user feedback to model user interests. Since user feedback can usually be directly learned by the algorithm, and there are many factors that can be directly learned about user satisfaction and engagement, recommendation systems usually model multiple aspects of user satisfaction or engagement in a multi-objective model-based manner.

[0003] This modeling method can transform input information into a shared vector by performing representation learning at the bottom layer of the network, and then input the shared vector into the sub-network corresponding to each task at the upper layer of the network to achieve prediction of multiple tasks. However, the prediction model obtained by this method is not accurate enough in different tasks, and the overall prediction performance of the model is poor. Summary of the invention

[0004] The embodiments of the present application provide an information processing method, apparatus, storage medium and computer equipment. The task weights can be modified through the personalized sub-network of each target task, thereby improving the performance of the model in multi-target prediction.

[0005] On the one hand, an embodiment of the present application provides an information processing method, which includes: obtaining object information, content information and background information, wherein the content information is information about the content to be recommended of the target object, and the background information is information about the scenario for recommending the content to be recommended; inputting the object information, content information and background information into a target prediction model to predict the recommendation score of the target content, and outputting the predicted score of the target content under different tasks; wherein the target prediction model is a sample target weight obtained by weight fusion based on sample personalized weights and sample task weights corresponding to different tasks, and is obtained by multi-task learning of a preset prediction network, wherein the sample personalized weight corresponding to each task is generated by a personalized subnetwork corresponding to each task based on the personalized characteristics of the object group, and the sample personalized weight corrects the sample task weight.

[0006] On the other hand, an embodiment of the present application also provides an information processing device, which includes: an information acquisition module, used to acquire object information, content information and background information, wherein the content information is information about the content to be recommended of the target object, and the background information is information about the scenario for recommending the content to be recommended; a score prediction model, used to input the object information, content information and background information into the target prediction model to predict the recommendation score of the target content, and output the predicted score of the target content under different tasks; wherein the target prediction model is a sample target weight obtained by weight fusion based on sample personalized weights and sample task weights corresponding to different tasks, and is obtained by multi-task learning of a preset prediction network, and the sample personalized weight corresponding to each task is generated by the personalized subnetwork corresponding to each task based on the personalized characteristics of the object group, and the sample personalized weight corrects the sample task weight.

[0007] On the other hand, an embodiment of the present application further provides a computer-readable storage medium, which stores a computer program, wherein the above-mentioned information processing method is executed when the computer program is executed by a processor.

[0008] On the other hand, an embodiment of the present application further provides a computer device, which includes a processor and a memory, wherein the memory stores a computer program, and the computer program executes the above-mentioned information processing method when called by the processor.

[0009] On the other hand, an embodiment of the present application also provides a computer program product, which includes a computer program stored in a storage medium; a processor of a computer device reads the computer program from the storage medium, and the processor executes the computer program, so that the computer device executes the steps in the above-mentioned computing engine determination method.

[0010] The present application provides an information processing method, which can obtain object information, content information and background information. Further, the object information, content information and background information are input into the target prediction model to predict the recommendation score of the target content, and the predicted score of the target content under different tasks is output. Among them, the target prediction model is a sample target weight obtained by weight fusion based on the sample personalized weights and sample task weights corresponding to different tasks, which is obtained by multi-task learning of the preset prediction network, and the sample personalized weight corresponding to each task is generated by the personalized sub-network corresponding to each task based on the personalized characteristics of the object group. Therefore, the task weight generated by the task gating is corrected by the personalized weight corresponding to each task, and the model's ability to model multiple targets is strengthened in the dimension of user personalization, so that the model can perform personalized learning for each task for different users, thereby enhancing the prediction accuracy of each task and improving the overall prediction performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0012] Figure 1 A schematic diagram of a system framework provided in an embodiment of the present application is shown.

[0013] Figure 2 A flow chart of an information processing method provided in an embodiment of the present application is shown.

[0014] Figure 3 The following is an architecture diagram of a video recommendation system provided in an embodiment of the present application.

[0015] Figure 4 A flow chart of another information processing method provided in an embodiment of the present application is shown.

[0016] Figure 5 An application scenario diagram of an information processing method provided by an embodiment of the present application is shown.

[0017] Figure 6 A network architecture diagram of a prediction network provided in an embodiment of the present application is shown.

[0018] Figure 7 It is a module block diagram of an information processing device provided in an embodiment of the present application.

[0019] Figure 8 It is a module block diagram of a computer device provided in an embodiment of the present application.

[0020] Fig. 9 It is a module block diagram of a computer-readable storage medium provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application.

[0022] In some processes described in the specification, claims and the above drawings, multiple steps appearing in a specific order are included, but it should be clearly understood that these steps may not be executed in the order in which they appear in this document or may be executed in parallel. The step numbers are only used to distinguish different steps, and the numbers themselves do not represent any execution order. In addition, the descriptions such as "first" and "second" in this document are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0023] In order to enable those skilled in the art to better understand the solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with 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 of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0024] It should be noted that in the specific implementation of this application, the object information, content information, background information and other related data involved, when applied to the specific products or technologies of the embodiments of this application, need to obtain user permission or consent, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of relevant countries and regions, and carry out subsequent data use and processing within the scope of authorization of laws, regulations and personal information subjects.

[0025] The introduction of recommender systems solves the problems of information overload and users’ lack of clear needs, and effectively helps users find content of interest. Recommender systems can usually be divided into recall, rough sorting, and fine sorting stages in terms of process. After the recall and rough sorting stages, the fine sorting stage requires more accurate sorting of the hundreds of candidate content (Items) screened out in the rough sorting stage. To this end, the recommender system generally weighs multiple business indicators in the fine sorting stage, adopts a multi-objective model structure, and sets corresponding target task process score predictions for different business indicators to finely sort the candidate content.

[0026] Generally, the structure of a multi-objective model mainly includes an upper network and a lower network. The lower network can concatenate features corresponding to different input information to obtain shared features. Further, the shared features are transmitted to the upper network to perform score prediction for tasks corresponding to multiple objectives, i.e., target tasks. For example, the multi-objective private expert network can enhance the modeling capability of multi-objective implicit crossover, and different target tasks can share the bottom input, i.e., shared features.

[0027] However, since the distribution of interest preferences of different users is significantly different, and the multi-objective private expert network cannot perceive such differences, the prediction accuracy of the multi-objective model for different target tasks is not high, and the prediction performance is poor overall. In order to solve the above problems, the inventor has proposed an information processing method provided by the embodiment of the present application after research. The following first introduces the system framework of the information processing method involved in the present application.

[0028] like Figure 1 As shown, the information processing method provided in the embodiment of the present application can be applied in the system 300, and the data acquisition device 310 is used to acquire training data. For the information processing method of the embodiment of the present application, the training data can be a training sample used for model training, and the training sample includes sample object information, sample content information, sample background information and scoring labels, wherein the scoring labels can be obtained by manual marking. After acquiring the training data, the data acquisition device 310 can store the training data in the database 320, and the training device 330 can train the target model 301 based on the training data maintained in the database 320.

[0029] Specifically, the training device 330 can train the preset neural network based on the input training data until the preset neural network meets the preset conditions and obtains the trained target model 301. Among them, the preset conditions can be: the total loss value of the target loss is less than the preset value, the total loss value of the target loss no longer changes, or the number of training times reaches the preset number of times, etc. The target model 301 can be used to implement the information processing method in the embodiment of the present application. The target model 301 in the embodiment of the present application can specifically be a deep neural network model, for example, a combination of an embedding layer representing learning and a multilayer perceptron (Multilayer Perceptron, MLP), etc., which is not limited here.

[0030] It should be noted that in actual application scenarios, the training data maintained in the database 320 does not necessarily all come from the data acquisition device 310, but may also be received from other devices. For example, the client device 360 ​​may also serve as a data acquisition terminal, and the acquired data is used as new training data and stored in the database 320. In addition, the training device 330 does not necessarily train the preset neural network based entirely on the training data maintained in the database 320, but may also train the preset neural network based on the training data obtained from the cloud or other devices. The above description should not be used as a limitation on the embodiments of the present application.

[0031] The target model 301 obtained by training the training device 330 can be applied to different systems or devices, such as Figure 1The execution device 340 shown may be a terminal, such as a mobile phone terminal, a tablet computer, a laptop computer, etc., or a server or a cloud, etc., but is not limited thereto. The execution device 340 may be used to interact with external devices for data. For example, a user (hereinafter collectively referred to as a target object) may use a client device 360 ​​to send input data to the execution device 340 via a network.

[0032] The input data in the embodiment of the present application may include: a request message sent by the client device 360. When the execution device 340 pre-processes the request message, or when the execution module 341 of the execution device 340 performs calculations and other related processing, the execution device 340 may call data, programs, etc. in the data storage system 350 for corresponding calculation processing, and store data and instructions such as processing results obtained by the calculation processing in the data storage system 350.

[0033] Finally, the execution device 340 can return the processing result, that is, the prediction result based on the target model 301, to the client device 360 ​​through the network, so that the target object can obtain the processing result on the client device 360. It is worth noting that the training device 330 can generate a corresponding target model 301 based on different training data for different goals or different tasks, and the corresponding target model 301 can be used to achieve the above goals or complete the above tasks, thereby providing the target object with the desired results.

[0034] For example, Figure 1 The system 300 shown may be a Client-Server (C / S) system architecture, the execution device 340 may be a cloud server deployed by a service provider, and the client device 360 ​​may be a laptop used by the target object. For example, the target object may use a video social software on a smart phone, and when the target object's authorization is obtained to use content information for video recommendation, the smart phone may send a request message to the cloud server through the network when the target object's software interface is reached.

[0035] Furthermore, when receiving the request message, the cloud server uses the target model 301, i.e., the target prediction model, to predict the scores of the candidate push content under different target tasks based on the object information, the content information of the candidate recommended content, and the background information. Then, the cloud server can calculate the recommended content with higher recommendation value among the candidate push content according to the predicted scores, and send the recommended content to the video social software for exposure (recommendation display).

[0036] It is worth noting that Figure 1It is only a schematic diagram of the architecture of a system provided in the embodiment of the present application. The architecture and application scenarios of the system described in the embodiment of the present application are intended to more clearly illustrate the technical solution of the embodiment of the present application, and do not constitute a limitation on the technical solution provided in the embodiment of the present application. Figure 1 The data storage system 350 is an external memory relative to the execution device 340. In other cases, the data storage system 350 can also be placed in the execution device 340. The above recommendation scenario can also be a product recommendation scenario. It is known to those skilled in the art that with the evolution of the system architecture and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to solving similar technical problems.

[0037] The information processing method of the present application involves artificial intelligence (AI) technology, which is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that the machines have the functions of perception, reasoning and decision-making.

[0038] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0039] Among them, deep learning (DL) is a sub-problem of machine learning, and its main purpose is to automatically learn effective feature representations from data. Through multi-layer feature conversion, the original data is transformed into a higher-level, more abstract representation. These learned representations can replace manually designed features, thereby avoiding "feature engineering". The abstract representation is further input into the prediction function to obtain the final result. For example, in an embodiment of the present application, the target prediction model can perform representation learning on the input object information, content information, and background information to obtain corresponding features.

[0040] See also Figure 2 , Figure 2 FIG. 1 is a flow chart of an information processing method provided by an embodiment of the present application. In a specific embodiment, the information processing method is applied to Figure 7 The information processing device 500 and the computer device 600 ( Figure 8 ).

[0041] The following will take a computer device as an example to illustrate the specific process of this embodiment. It can be understood that the computer device used in this embodiment can be a server or a terminal, etc. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, blockchain, and big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., but is not limited to this. The information processing method can specifically include the following steps:

[0042] Step 110: Obtain object information, content information and background information.

[0043] Considering that with the expansion of recommendation business scenarios, recommendation systems often need to weigh multiple business indicators at the same time. Therefore, a modeling method for a multi-objective model is proposed. That is, for the target object, the score of the target content (the content to be recommended) under different target tasks is predicted. The object information is the relevant information of the user to be recommended (that is, the target object).

[0044] The content information may refer to the relevant information of the target object's content (Item) to be recommended, which may include basic information, statistical information, and cross information. The content to be recommended may include commodities, advertisements, videos, and other types of content. Taking the video type as an example, the basic information may include the content identifier (Item ID), category, listing time, definition, physical duration, etc. of the content to be recommended.

[0045] Statistical information may include the video's exposure (Page View, PV), number of visits (Visit View, VV), number of unique visitors (Unique Visitor, UV) and completion rate in the past 14 days, 7 days or 3 days. Cross-information may be cross-information between recommended content and target objects, for example, statistical features of recommended content on target objects of different genders and ages.

[0046] The background information may refer to information related to the scenario in which the content is recommended, and may include the recommendation time, such as the day of the week or time of the day when the content is currently recommended, or the recommendation time may be a period set according to business needs, such as the off-season or peak season for recommendation. The background information may also include the network type, operating system, and client version used for the recommendation.

[0047] Optionally, the background information may also be a recommendation scenario, including a flow identifier or a layout identifier. The flow identifier may be used to indicate in which specific application the recommendation is made, such as a social network or an open application platform. The layout identifier is used to indicate the form of the recommendation, such as a start screen recommendation of an application or a horizontal screen recommendation in a video stream in a video player.

[0048] As an implementation method, the target object's object information can be obtained in response to a prediction request of the target object, and the content information of each candidate content can be obtained from the candidate content set and the background information can be obtained. For example, when the target object is browsing a web page, a prediction request can be generated. Further, in response to the prediction request, the target object's object information can be obtained from the service log, and the content information of each candidate content can be obtained from the candidate content set in the content database. Further, the time when the target object browses the web page and the recommended layout in the web page can be obtained.

[0049] It should be noted that when obtaining background information and target object information, it is necessary to follow the principles of legality, legitimacy, necessity and good faith in accordance with the requirements of laws and regulations, and only collect information necessary to achieve content recommendation. For example, before making content recommendations, the recommendation function management interface can be displayed to the target object in the form of a pop-up window so that the target object can determine whether to turn on the recommendation function. When the target object turns on the recommendation function, the target object's basic attribute information (object ID and name, etc.) and behavior information (clicks and sharing, etc.) can be obtained.

[0050] Step 120: Input the object information, content information and background information into the target prediction model to predict the recommendation score of the target content, and output the predicted score of the target content under different tasks.

[0051] In the recommendation scenario, in order to predict different business indicators at the same time, the existing technology sets corresponding target tasks for different business indicators, such as video viewing rate, sharing rate, and comment rate, and then each target task can train a multi-target model based on shared features and its own corresponding sub-network. However, this approach ignores the personalized differences of different target objects, that is, each target task does not take into account the personalized distribution of different target objects during execution.

[0052] For example, for videos recommended by the system, at the level of activity, the click behavior and viewing time of the target object can show a large consistency, while the viewing time and interactive behavior show a large difference. Usually, high interaction corresponds to short viewing time, and high viewing time corresponds to low interaction. At the level of basic attributes of the target object, young objects correspond to high interaction, and old objects correspond to low interaction.

[0053] When a multi-objective model is trained based on shared features, the prediction performance of the trained multi-objective model for different target tasks may be such that the prediction accuracy of some tasks is high, while the prediction accuracy of other tasks is low. As a result, when the trained multi-objective model predicts different tasks, the prediction performance of each task is inaccurate and unstable, and the overall prediction performance of the model is poor. To this end, the present application creatively proposes a target prediction model, which calculates the prediction score of the target content under different tasks based on the personalized distribution of different target objects.

[0054] The target content refers to each candidate content in the candidate content set that is predicted for rating using the target prediction model. The target prediction model is a sample target weight obtained by weight fusion based on the sample personalized weights corresponding to different tasks and the sample task weights. It is obtained by multi-task learning of the preset prediction network. The sample personalized weight corresponding to each task is generated by the personalized subnetwork corresponding to each task based on the personalized characteristics of the object group.

[0055] In this way, the task weights generated by task gating are corrected through the personalized weights corresponding to each task, and the model's ability to differentiated modeling of multiple targets is enhanced based on the personalized dimensions of the target objects. As a result, the model can capture personalized information of different target objects during the execution of multi-target tasks, thereby improving the accuracy and stability of the target prediction model's prediction performance under multi-target tasks, and effectively improving the overall prediction performance of the model.

[0056] As an implementation method, object information, content information and background information can be input into a target prediction model, and then for the target object, each predicted candidate content can be predicted to have a corresponding prediction score under different tasks. Specifically, object information, content information and background information can be input into a target prediction model for feature extraction to obtain common features. The target prediction model can include a shared subnetwork and a task subnetwork corresponding to each task.

[0057] Furthermore, the common features can be transmitted to the shared subnetwork and the task subnetwork corresponding to each task for feature crossover to obtain the shared features corresponding to the shared subnetwork and the task features corresponding to each task subnetwork, and through the task gating corresponding to each task, the task weight corresponding to each task is generated based on the common features. Furthermore, through the personalized subnetwork corresponding to each task, the personalized weight corresponding to each task can be generated based on the personalized features of the object group, and the score prediction is performed based on the task weight corresponding to each task and the personalized weight to obtain the predicted score of each candidate content under different tasks.

[0058] See also Figure 3 , Figure 3 The architecture diagram of a video recommendation system is shown in FIG. The system can recommend videos to target users based on the feedback of different target objects. Figure 3 As shown, in the video recommendation system, there are multiple goals to model the behaviors of different target objects. For example, multiple goals can be determined according to business indicators, including video completion rate (VCR), view through rate (VTR), share rate (SHR), and comment rate (CMR). To this end, the tasks corresponding to the multiple goals determined by the business indicators can be the task of predicting the view completion rate, the task of predicting the view through rate, the task of predicting the share rate, and the task of predicting the comment rate.

[0059] During offline training, the model can be trained based on the target object behavior extracted from the service log to obtain a target prediction model. At each online request, the prediction task of each target in the target prediction model can output a corresponding prediction score, and then these prediction scores can be combined into a comprehensive score, and the target video with a high comprehensive score ranking can be recommended to the target object.

[0060] Specifically, the target prediction model can be used to perform representation learning based on object information, content information and background information of each candidate video, and obtain corresponding object features, content features and background features, where each feature can be in the form of a vector in the embedding space. The object features, content features and background features are concatenated to obtain common features. The common features are respectively transmitted to the shared subnetwork and the task subnetwork corresponding to each prediction task for feature crossover, and the shared features corresponding to the shared subnetwork and the task features corresponding to each task subnetwork are obtained.

[0061] Furthermore, the task weight corresponding to each prediction task can be generated based on the common features through the task gating corresponding to each prediction task, and the personalized weight corresponding to each prediction task can be generated based on the personalized features of the object group through the personalized sub-network corresponding to each prediction task. Furthermore, the score prediction can be performed based on the task weight corresponding to each prediction task and the personalized weight to obtain the predicted score of each candidate video under different target tasks. Where n is the number of target tasks.

[0062] Furthermore, a comprehensive score can be obtained by weighted multiplication of these predicted scores, and all candidate videos can be sorted according to the comprehensive score, and then the candidate videos corresponding to the top-k comprehensive scores can be selected as target videos to be recommended to the target object. For example, the target object can be recommended to the interface of the video social software that the target object is browsing.

[0063] In the embodiment of the present application, object information, content information and background information can be obtained. Furthermore, the object information, content information and background information can be input into the target prediction model to output the predicted score of the target content under different tasks. Among them, the target prediction model is a target weight obtained by weight fusion based on the personalized weights corresponding to different tasks and the task weights, which is obtained by multi-task learning of the preset prediction network, and the personalized weight corresponding to each task is generated by the personalized sub-network corresponding to each task based on the personalized characteristics of the object group. Therefore, the task weight generated by the task gating is corrected by the personalized weight corresponding to each task, which can enhance the model's ability to correspond to multi-objective modeling in the dimension of user personalization, so that the model can perform personalized learning for each task for different users, enhance the prediction accuracy of each task, and thus improve the overall prediction performance of the model.

[0064] In combination with the method described in the above embodiment, the method will be further described in detail with examples below.

[0065] The information processing method provided in this embodiment specifically relates to the technology in the field of artificial intelligence. Figure 4 The process shown combines Figure 4 The application scenario shown in the figure is described in detail. The computer device can be a server or a terminal device, etc. Figure 5 , Figure 5 Another information processing method provided by the embodiment of the present application is shown. In a specific embodiment, the information processing method can be applied to Figure 5 The video shown is recommended in the scenario.

[0066] The video service provider provides a service end, which includes a cloud training server 410 and a cloud execution server 430. The cloud training server 410 can be used to train a target prediction model for video recommendation, and the cloud execution server 430 is used to deploy the target prediction model trained on the cloud training server 410, and predict the corresponding prediction score of each candidate video in different tasks based on the target prediction model, video information (that is, content information), object information and background information.

[0067] Furthermore, the cloud execution server 430 can be used to predict the scores of each candidate video in different tasks, determine the comprehensive scores of each candidate video, and then determine the target video according to the comprehensive scores, and expose the target video on the client. For example, the client can be a desktop computer 420, and the cloud execution server 430 can put the target video on the video social software 421 installed on the desktop computer 420 for display.

[0068] It should be noted that Figure 5 It is only an application scenario provided by the embodiment of the present application. The application scenario described in the embodiment of the present application is to more clearly illustrate the technical solution of the embodiment of the present application, and does not constitute a limitation on the technical solution provided by the embodiment of the present application. For example, in other cases, the desktop computer 420 can also deploy a target prediction model, so that the entire process of video recommendation can be completed directly on the desktop computer 420. The target prediction model can also be used in application scenarios of product recommendation or advertising. It is known to those of ordinary skill in the art that with the evolution of system architecture and the emergence of new application scenarios, the technical solution provided in the embodiment of the present application is also applicable to solving similar technical problems. The information processing method may specifically include the following steps:

[0069] Step 210: The computer device obtains training samples.

[0070] The training samples include sample object information, sample content information, sample background information and rating labels. Considering the heterogeneous sample space caused by the continuous behavior of the target object. For example, a user can only share or comment on a video after clicking on it, so different tasks will have different sample spaces for training.

[0071] To this end, the training samples may include at least one of an exposure sample (ImpressionSample), a play sample (PlaySample) and a click sample (ClickSample). The exposure sample can be used for task training to predict VTR and CTR (ClickThroughRate), the play sample can be used for task training to predict VCR, and the click sample can be used for training to predict SHR and CMR tasks.

[0072] Each type of training sample has corresponding positive and negative samples. Specifically, the training samples can be labeled by the behavioral feedback of the target object to obtain the score label corresponding to each training sample. For example, for the prediction of SHR and CMR, model training can be based on the playback hypothesis. The target object's sharing or commenting behavior represents positive feedback, indicating that the video corresponding to the training sample is attractive to the target object. The score label corresponding to the training sample can be marked as 1. Conversely, playback without sharing or commenting represents negative feedback, and the score label can be marked as 0.

[0073] As an implementation method, the cloud training server 410 can obtain a training set D={(x i ,y i )|i=1,2,…,N},where x irepresents the i-th training sample, y i Represents the training sample x i The corresponding score label, N is the number of training samples.

[0074] It is worth noting that the information processing method provided in the embodiment of the present application includes training of a preset prediction network. The training process can be performed based on the acquired training set D. Subsequently, each time a prediction score is required, the target prediction model obtained by training can be used for direct calculation without the need to perform a prediction score each time and repeatedly perform training based on the training set D.

[0075] Step 220: The computer device inputs the sample object information, sample content information and sample background information into the prediction network for feature extraction to obtain corresponding sample object features, sample content features and sample background features.

[0076] See also Figure 6 , Figure 6 A network architecture diagram of a prediction network is shown. Figure 6 As shown, the subnetwork of the prediction network may include a tower subnetwork, a task subnetwork, a shared subnetwork, a task gating, a personalized subnetwork, and an embedding layer. The computer device may input the sample object information, sample content information, and sample background information of the training sample into the prediction network, and perform feature extraction through the embedding layer of the prediction network to obtain the corresponding sample push features, sample object features, and sample scene features. The task gating may be modified based on the personalized subnetwork.

[0077] The tower subnetwork can be a multi-layer network, and its width and depth are hyperparameters. Specifically, the shared subnetwork is responsible for learning shared knowledge, while the knowledge of different target tasks is learned by the specific task subnetwork. Each tower subnetwork can learn knowledge from the shared subnetwork and its own specific task subnetwork. The number of tower subnetworks and task subnetworks can be determined by the number of tasks, and the number of shared subnetworks can be set according to actual experimental experience and application needs.

[0078] In the embodiment of the present application, the training score can be predicted based on the completion rate and the browsing rate. Correspondingly, the prediction network is used to perform two tasks, a first task of predicting the completion rate score and a second task of predicting the browsing rate score. This embodiment only takes the first task and the second task as examples, and the prediction network can also be used to perform multiple tasks of other types.

[0079] For example, the cloud training server 410 may send a training sample, which may include sample content information x item , sample object information x user and sample background information xcontext Input to the prediction network. Then, the corresponding sample content feature emb is obtained through the embedding layer representation learning of the prediction network. item , sample object features emb user and sample background features emb context .

[0080] Step 230: The computer device uses a prediction network to perform score prediction based on sample object features, sample content features, and sample background features to obtain training scores for training samples corresponding to different tasks.

[0081] Since a target prediction model for performing a specific target task can be trained specifically based on different types of training samples, the training score may include at least one of a viewing completion rate, a browsing rate, a sharing rate, and a comment rate.

[0082] As an implementation manner, the computer device performs score prediction based on sample object features, sample content features, and sample background features through a prediction network to obtain training scores of training samples corresponding to different tasks, which may include:

[0083] (1) The computer device uses a prediction network to perform vector concatenation on sample object features, sample content features, and sample background features to obtain sample common features.

[0084] Exemplarily, the cloud training server 410 obtains the sample content feature emb item , sample object features emb user and sample background features emb context When , the three can be concatenated to obtain the sample common feature emb com .

[0085] For example, the sample content feature emb item =(0.585, 0.168, 2.785), sample object feature emb user =(4.565, 1.136, 0.113) and sample background features emb context =(0.625, 0.498, 1.374). By concatenating the sample content features, sample object features, and sample background features, we can get the sample universal feature emb com =(0.585, 0.168, 2.785, 4.565, 1.136, 0.113, 0.625, 0.498, 1.374).

[0086] (2) The computer device transmits the common features of the samples to the shared subnetwork and the task subnetwork corresponding to each task for feature cross-pollination, thereby obtaining the sample shared features corresponding to the shared subnetwork and the sample task features corresponding to each task subnetwork.

[0087] Considering the indiscriminate sharing of common features based on samples, the prediction network will fail to capture the correlation between complex tasks and may bring harmful noise to some tasks. Therefore, by explicitly separating the shared subnetwork and the task subnetwork, more efficient knowledge transfer can be achieved between complex related tasks to avoid harmful parameter interference.

[0088] Exemplarily, the cloud training server 410 can input the sample common features into the shared subnetwork and the task subnetwork corresponding to each task respectively, and then the shared subnetwork and the task subnetwork corresponding to each task can perform feature crossover based on the sample common features to obtain the sample shared features corresponding to the shared subnetwork and the sample task features corresponding to each task subnetwork.

[0089] (3) The computer device performs feature splicing based on the sample shared features and each sample task features to obtain a sample selection matrix.

[0090] In an embodiment of the present application, the sample selection matrix is ​​a matrix obtained by concatenating the high-order outputs of the shared sub-network and the task sub-network, and each column of the sample selection matrix represents a high-order output vector of a sub-network.

[0091] For example, the cloud training server 410 may concatenate the high-order outputs of the shared subnetwork and the task subnetwork to obtain the sample selection matrix S k (x):

[0092]

[0093] Among them, k is the task ordinal number, e s is the high-order output of the shared subnetwork, e k is the high-order output of the task subnetwork corresponding to the kth task, m s is the number of shared subnetworks, m p is the number of task sub-networks, and T is the transpose of the high-order output.

[0094] (4) The computer device performs weight calculation based on the task gating and personalized sub-network corresponding to each task to obtain the sample target weight corresponding to each task.

[0095] Taking into account the personalized differences of the target objects in different tasks, for the same target object, when the outputs of the shared subnetwork and the task subnetwork are weighted combined in each task, a corresponding personalized subnetwork is set for each task, and then the corresponding task gating is corrected using the personalized subnetwork. In this way, the model's ability to model multi-target differentiation is enhanced in the personalized dimension of the target object, thereby improving the overall prediction effect of the trained target prediction model.

[0096] In some embodiments, the computer device performs weight calculation based on the task gating and personalized sub-network corresponding to each task, and the step of obtaining the sample target weight corresponding to each task may also include:

[0097] (4.1) The common features of the samples are transferred to the task gate corresponding to each task for weighted calculation to obtain the sample task weight corresponding to each task.

[0098] Among them, task gating can better capture task differences and effectively optimize multiple target tasks. The structure of task gating can be based on a single-layer feedforward network with SoftMax as the activation function.

[0099] For example, the cloud training server 410 can transmit the sample common features to the task gate corresponding to each task for weighted calculation to obtain the sample task weight G corresponding to each task. k (x), the calculation formula can be shown as follows:

[0100] G k (x) = Softmax(W k x) (1)

[0101] in, is the parameter matrix, D is the dimension of the common features of the samples, and x is the common features of the samples.

[0102] (4.2) Through the personalized sub-network corresponding to each task, the sample personalized weight corresponding to each task is generated based on the personalized characteristics of the object group.

[0103] Under multi-target conditions, different target objects have large individual differences, which also makes the samples sparse and causes extremely unbalanced sample distribution. The intuitive and naive solution is to use a separate model to learn this type of distribution, forcing the model to pay attention to this personalized feature.

[0104] To this end, this application can learn the corresponding model for each target object, without being affected by other samples, and can be completely personalized based on the target object itself. Based on this idea, a personalized sub-network is introduced for each task, so that the model can respond to each target object in a personalized manner at the multi-target model structure level, realizing the personalization of the model structure.

[0105] However, considering that the corresponding model is learned for each target object, although it is a personalized recommendation model, it is not feasible at the implementation level due to the large number of target objects in the recommendation scenario. In addition, a single user can be regarded as a combination of different user groups (that is, object groups). For example, a user can be regarded as a combination of active users or inactive users, that is, active groups or inactive groups. For example, the object groups of new users, the object groups of old users, the object groups of returning users, etc., and the interest preferences of these object groups can be modeled.

[0106] Among them, the personalized features at least include the activity features of the group to which the target object belongs (that is, the object group), for example, the activity features of watching 30 seconds, specifically, the behavior sequence length, total number of play visits, number of likes, etc. of the target object corresponding to the video with a play time of more than 30 seconds, which reflect the activity level of the target object. If the population division based on the basic attributes of the target object is considered, some basic attribute features of the target object (such as gender, age, etc.) can also be added, or features such as the target object ID can be directly added. The vector representation of the personalized features can be calculated in advance by other models.

[0107] See also Figure 6 , the network structure of the personalized subnetwork can be a lightweight deep neural network (Deep Neural Network, DNN). It can also be a simple softmax transformation, which is not limited here. Optionally, the personalized subnetwork can be a gated network composed of multiple fully connected layers, and the personalized subnetwork can also be a network that performs linear transformation on the input based on the transformation matrix. For example, a personalized subnetwork composed of three fully connected layers inputs the personalized features into the first fully connected layer of the personalized subnetwork to obtain the first output, and then inputs the first output into the second fully connected layer to obtain the second output, and then inputs the second output into the third fully connected layer, that is, the output layer, to obtain the sample personalized weight.

[0108] For example, the cloud training server 410 may obtain the personalized feature x corresponding to each target task. pc , and through the personalized sub-network corresponding to each task, based on the personalized features x of the object group pc Generate sample personalized weights p for each task k (x pc ).

[0109] (4.3) Based on the sample task weight and sample personalized weight corresponding to each task, a bitwise multiplication operation is performed to obtain the sample target weight corresponding to each task.

[0110] For example, the cloud training server 410 can perform an element-wise product operation based on the sample task weight and the sample personalized weight corresponding to each task, that is, "⊙", to obtain the sample target weight Q corresponding to each task. k , the formula can be shown as follows:

[0111] Q k =G k (x)⊙p k (x pc ) (2)

[0112] (5) The computer device predicts the score based on the sample target weight corresponding to each task and the sample selection matrix to obtain the training scores of the training samples corresponding to different tasks.

[0113] For example, the cloud training server 410 can perform score prediction based on the sample target weight corresponding to each task and the sample selection matrix to obtain the training scores of the training samples corresponding to different tasks. The formula can be shown as follows:

[0114]

[0115] R k (x) = S k (x)·[Q k ] T =S k (x)·[G k (x)⊙p k (x pc )] T (4)

[0116] Among them, H k (·) is the tower sub-network corresponding to the kth task.

[0117] Step 240: The computer device performs a cross entropy operation based on the training scores and the score labels to determine the target loss corresponding to the prediction network.

[0118] (1) The computer device performs a cross entropy operation based on the training score and score label corresponding to each task to determine the cross entropy loss corresponding to each task.

[0119] Different weights can be set for different target tasks, and then the cross entropy losses of all target tasks are weighted according to the weights to obtain the target loss corresponding to the prediction network.

[0120] For example, the cloud training server 410 can calculate the training score corresponding to each task. and rating tags Perform cross entropy calculation to determine the cross entropy loss corresponding to each task Among them, i is the ordinal number of the training sample.

[0121] (2) The computer device determines the task loss corresponding to each task based on the spatial parameters and cross entropy loss of the training samples.

[0122] In order to solve the problem of inconsistent sample spaces, when calculating the loss of each task, we need to merge the samples with the same sample space and ignore the samples that are not in their own sample space, that is, different tasks still use the samples in their respective sample spaces.

[0123] For example, the cloud training server 410 may determine the task loss L corresponding to each task based on the spatial parameters and cross entropy loss of the training samples. k (θ k ,θ s ), the formula can be shown as follows:

[0124]

[0125] in, is the spatial parameter of the i-th sample under the k-th task. δ takes the value of 0 or 1. θ s Represents shared parameters. k Represents the parameters to be trained for the kth task.

[0126] (3) The computer device determines the target loss corresponding to the prediction network based on the loss weight corresponding to each task and the task loss.

[0127] Exemplarily, the cloud training server 410 may determine the target loss corresponding to the prediction network based on the loss weight corresponding to each task and the task loss. The formula may be as follows:

[0128]

[0129] Where K is the number of target tasks. k ,ω k Denote the loss function and loss weight of the k-th task, respectively. Considering that the model performance of multi-task learning is very sensitive to the choice of loss weights during training, it determines the relative importance of each task to the joint loss (i.e., the target loss). In practice, it can be observed that each task may have different importance at different stages of training. Therefore, the loss weight of each task can be regarded as a dynamic weight rather than a static weight. The dynamic update of the loss weight can be dynamically adjusted during the training process to update the actual experimental needs.

[0130] Step 250: The computer device performs iterative training of multi-task learning of different tasks on the prediction network based on the target loss until the prediction network meets preset conditions and obtains a target prediction model.

[0131] For example, the cloud training server 410 may target the loss L(θ1, . . . , θ K ,θ s ) Iteratively train the prediction network for multi-task learning of different target tasks until the prediction network meets the preset conditions and obtains the target prediction model. The preset conditions may be: the total loss value of the target loss is less than the preset value, the total loss value of the target loss no longer changes, or the number of training times reaches the preset number of times. Optionally, an optimizer may be used to optimize the target loss, and the learning rate, batch size during training, and epoch of training may be set based on experimental experience.

[0132] After the cloud training server 410 obtains the target prediction model through training, the target prediction model may be deployed on the cloud execution server 430 to provide video services.

[0133] Step 260: The computer device obtains object information, content information, and background information.

[0134] Exemplarily, the cloud execution server 430 can obtain the object information of the target object in response to the prediction request of the target object, and obtain the content information of each candidate content from the candidate content set and the background information. For example, when the target object browses a web page, a prediction request can be generated. Further, in response to the prediction request, the object information of the target object can be obtained from the service log, and the content information of each candidate content can be obtained from the candidate content set in the content database. Further, the time when the target object browses the web page and the recommended layout in the web page can be obtained.

[0135] Step 270: The computer device inputs the object information, content information and background information into the target prediction model to predict the recommendation score of the target content, and outputs the predicted score of the target content under different tasks.

[0136] As an implementation method, the cloud execution server 430 can input object information, content information and background information into the target prediction model for feature extraction to obtain common features. The common features are respectively transmitted to the shared subnetwork and the task subnetwork corresponding to each task for feature crossover to obtain the shared features corresponding to the shared subnetwork and the task features corresponding to each task subnetwork.

[0137] Furthermore, the cloud execution server 430 can generate a task weight corresponding to each task based on the shared features and each task feature through the task gating corresponding to each task, and generate a personalized weight corresponding to each task based on the personalized features of the object group through the personalized sub-network corresponding to each task. Then, the score prediction is performed based on the task weight corresponding to each task and the personalized weight to obtain the predicted score of each candidate content under different tasks.

[0138] Specifically, the completion rate and the browsing rate can be used as the predicted training score. Correspondingly, the target prediction model is used to perform two tasks, a first task of predicting the completion rate score and a second task of predicting the browsing rate score. The cloud execution server 430 can determine the first target weight according to the task weight and the personalized weight corresponding to the first task, and determine the second target weight according to the task weight and the personalized weight corresponding to the second task.

[0139] Furthermore, the cloud execution server 430 can perform a score prediction based on the first target weight and the selection matrix corresponding to the first task to obtain a first predicted score for each candidate content under the first task, and perform a score prediction based on the second target weight and the selection matrix corresponding to the second task to obtain a second predicted score for each candidate content under the second task.

[0140] Furthermore, the cloud execution server 430 can determine a comprehensive score based on the first prediction score and the second prediction score, and sort each candidate content in descending order according to the comprehensive score to obtain a content sequence. Further, from the content sequence, determine a preset number of candidate contents corresponding to the preceding comprehensive scores as target content for recommendation.

[0141] Further, the cloud execution server 430 may determine multiple target contents and display the multiple target contents on the interface displayed by the video social software 421. For example, see Figure 5 ,While the target object is watching the video, each target video can be displayed in sequence according to the rating in the waiting to be played column.

[0142] In the embodiment of the present application, a training sample can be obtained, the training sample includes sample object information, sample content information, sample background information and a score label, and the sample object information, sample content information and sample background information are input into the prediction network for feature extraction to obtain the corresponding sample object features, sample content features and sample background features. Further, through the prediction network, score prediction is performed based on the sample object features, sample content features and sample background features to obtain the training scores of the training samples corresponding to different tasks, and a cross entropy operation is performed based on the training scores and score labels to determine the target loss corresponding to the prediction network, and then the prediction network is iteratively trained for multi-task learning of different tasks based on the target loss until the prediction network meets the preset conditions and the target prediction model is obtained.

[0143] Furthermore, object information, content information and background information can be obtained, and further, the object information, content information and background information are input into the target prediction model to output the prediction score of the target content under different tasks. Then, through the personalized weight corresponding to each task, the task weight generated by the task gating is corrected, and the ability of the model to correspond to multi-objective modeling is enhanced in the dimension of user personalization, so that the model can perform personalized learning for each task for different users, enhance the prediction accuracy of each task, and thus improve the overall prediction performance of the model.

[0144] See also Figure 8 , which shows a structural block diagram of an information processing device 500 provided in an embodiment of the present application. The information processing device 500 includes: an information acquisition module 510, which is used to acquire object information, content information and background information, wherein the content information is information about the content to be recommended of the target object, and the background information is information about the scene for recommending the content to be recommended; a score prediction model 520, which is used to input the object information, content information and background information into the target prediction model to predict the recommendation score of the target content, and output the predicted score of the target content under different tasks; wherein the target prediction model is a sample target weight obtained by weight fusion based on the sample personalized weights and sample task weights corresponding to different tasks, obtained by multi-task learning of the preset prediction network, the sample personalized weight corresponding to each task is generated by the personalized sub-network corresponding to each task based on the personalized characteristics of the object group, and the sample personalized weight modifies the sample task weight.

[0145] In some embodiments, the information processing device 500 may further include: a sample acquisition module, a feature extraction module, a score calculation module, a loss determination module, and a model training module. The sample acquisition module is used to obtain training samples, and the training samples include sample object information, sample content information, sample background information, and score labels; the feature extraction module is used to input the sample object information, sample content information, and sample background information into the prediction network for feature extraction, and obtain the corresponding sample object features, sample content features, and sample background features; the score calculation module is used to perform score prediction based on the sample object features, sample content features, and sample background features through the prediction network, and obtain the training scores of the training samples corresponding to different tasks; the loss determination module is used to perform cross entropy operations based on the training scores and score labels to determine the target loss corresponding to the prediction network; the model training module is used to perform iterative training of multi-task learning of different tasks on the prediction network based on the target loss until the prediction network meets the preset conditions and obtains the target prediction model.

[0146] In some embodiments, the score calculation module may include a vector splicing unit, a feature cross unit, a feature fusion unit, a weight determination unit, and a score prediction unit. The vector splicing unit is used to perform vector splicing on sample object features, sample content features, and sample background features through a prediction network to obtain sample common features; the feature cross unit is used to transmit the sample common features to the shared subnetwork and the task subnetwork corresponding to each task to perform feature cross, to obtain sample shared features corresponding to the shared subnetwork and sample task features corresponding to each task subnetwork; the feature splicing unit is used to perform feature splicing based on the sample shared features and each sample task feature to obtain a sample selection matrix; the weight determination unit is used to perform weight calculation on the task gating and personalized subnetwork corresponding to each task to obtain the sample target weight corresponding to each task; the score prediction unit is used to perform score prediction based on the sample target weight corresponding to each task and the sample selection matrix to obtain the training scores of the training samples corresponding to different tasks.

[0147] In some embodiments, the weight determination unit can be specifically used to: transmit the common features of the samples to the task gate corresponding to each task for weighted calculation to obtain the sample task weight corresponding to each task; generate the sample personalized weight corresponding to each task based on the personalized features of the object group through the personalized sub-network corresponding to each task, and the personalized features at least include the activity features of the object group; perform a bitwise multiplication operation based on the sample task weight and the sample personalized weight corresponding to each task to obtain the sample target weight corresponding to each task.

[0148] In some embodiments, the loss determination module can be specifically used to: perform cross entropy calculations based on the training scores and score labels corresponding to each task to determine the cross entropy loss corresponding to each task; determine the task loss corresponding to each task based on the spatial parameters of the training samples and the cross entropy loss; determine the target loss corresponding to the prediction network based on the loss weight and task loss corresponding to each task.

[0149] In some embodiments, the score prediction model 520 may include a feature extraction unit, a feature crossover unit, a task gating unit, a personalized sub-network unit, and a score calculation unit. The feature extraction unit is used to input object information, content information, and background information into the target prediction model for feature extraction to obtain common features; the feature crossover unit is used to transmit the common features to the shared sub-network and the task sub-network corresponding to each task for feature crossover to obtain the shared features corresponding to the shared sub-network and the task features corresponding to each task sub-network; the task gating unit is used to generate the task weight corresponding to each task based on the common features through the task gating corresponding to each task; the personalized sub-network unit is used to generate the personalized weight corresponding to each task based on the personalized features of the object group through the personalized sub-network corresponding to each task; the score calculation unit is used to perform score prediction according to the task weight corresponding to each task and the personalized weight to obtain the predicted score of each candidate content under different tasks.

[0150] In some embodiments, the task includes a first task and a second task, and the score calculation unit can be specifically used to: determine a first target weight according to the task weight and personalized weight corresponding to the first task; determine a second target weight according to the task weight and personalized weight corresponding to the second task; perform score prediction based on the first target weight and the selection matrix corresponding to the first task to obtain a first predicted score for each candidate content under the first task. Perform score prediction based on the second target weight and the selection matrix corresponding to the second task to obtain a second predicted score for each candidate content under the second task.

[0151] In some embodiments, the rating prediction model 520 can also be specifically used to: determine a comprehensive rating based on a first predicted rating and a second predicted rating; sort each candidate content in descending order according to the comprehensive rating to obtain a content sequence; and determine, from the content sequence, a preset number of candidate contents corresponding to the preceding comprehensive ratings for recommendation as target content.

[0152] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here.

[0153] In several embodiments provided in the present application, the coupling between modules may be electrical, mechanical or other forms of coupling.

[0154] In addition, each functional module in each embodiment of the present application can be integrated into a processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The above integrated modules can be implemented in the form of hardware or software functional modules.

[0155] The solution provided by the present application can obtain object information, content information and background information. Further, the object information, content information and background information are input into the target prediction model to output the predicted score of the target content under different tasks. Among them, the target prediction model is a target weight obtained by weight fusion based on the personalized weights corresponding to different tasks and the task weights, and is obtained by multi-task learning of the preset prediction network. The personalized weight corresponding to each task is generated by the personalized sub-network corresponding to each task based on the personalized characteristics of the object group. Therefore, the task weight generated by the task gating is corrected by the personalized weight corresponding to each task, and the ability of the model to correspond to multi-objective modeling is strengthened in the dimension of user personalization, so that the model can perform personalized learning for each task for different users, enhance the prediction accuracy of each task, and thus improve the overall prediction performance of the model.

[0156] like Figure 8 As shown, the embodiment of the present application further provides a computer device 600, which includes a processor 610, a memory 620, a power supply 630 and an input unit 640. The memory 620 stores a computer program. When the computer program is called by the processor 610, the various method steps provided in the above embodiment can be implemented. Those skilled in the art can understand that the structure of the computer device shown in the figure does not constitute a limitation on the computer device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. Among them:

[0157] The processor 610 may include one or more processing cores. The processor 610 uses various interfaces and lines to connect various parts of the entire battery management system, and by running or executing instructions, programs, instruction sets or program sets stored in the memory 620, calls the data stored in the memory 620, executes various functions of the battery management system and processes data, and executes various functions and processes data of the computer device, thereby controlling the computer device as a whole. Optionally, the processor 610 can be implemented in at least one hardware form of digital signal processing (DSP), field programmable gate array (FPGA), and programmable logic array (PLA). The processor 610 can integrate one or more combinations of a central processing unit 610 (CPU), a graphics processing unit 610 (GPU), and a modem. Among them, the CPU mainly processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing display content; and the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 610, but may be implemented separately through a communication chip.

[0158] The memory 620 may include a random access memory 620 (Random Access Memory, RAM), and may also include a read-only memory 620 (Read-Only Memory). The memory 620 may be used to store instructions, programs, instruction sets or program sets. The memory 620 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc. The data storage area may also store data (such as a phone book and audio and video data) created by the computer device during use. Accordingly, the memory 620 may also include a memory controller to provide the processor 610 with access to the memory 620.

[0159] The power supply 630 can be logically connected to the processor 610 through a power management system, so as to manage charging, discharging, power consumption and other functions through the power management system. The power supply 630 can also include one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, power status indicators and other arbitrary components.

[0160] The input unit 640 can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0161] Although not shown, the computer device 600 may also include a display unit, etc., which will not be described in detail herein. Specifically in this embodiment, the processor 610 in the computer device will load the executable files corresponding to the processes of one or more computer programs into the memory 620 according to the following instructions, and the processor 610 will run the data stored in the memory 620, such as the phone book and audio and video data, to implement the various method steps provided in the aforementioned embodiments.

[0162] like Fig. 9 As shown, the embodiment of the present application further provides a computer-readable storage medium 700, in which a computer program 710 is stored. The computer program 710 can be called by a processor to execute various method steps provided in the embodiment of the present application.

[0163] The computer readable storage medium may be an electronic memory such as a flash memory, an EEPROM (Electrically Erasable Programmable Read-Only Memory), an EPROM, a hard disk, or a ROM. Optionally, the computer readable storage medium includes a non-volatile computer readable storage medium (Non-Transitory Computer-Readable Storage Medium). The computer readable storage medium 700 has storage space for a computer program that executes any of the method steps in the above embodiments. These computer programs can be read from or written to one or more computer program products. The computer program can be compressed in an appropriate form.

[0164] According to one aspect of the present application, a computer program product is provided, the computer program product comprising a computer program, the computer program being stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device executes various method steps provided in the above embodiments.

[0165] The above are only preferred embodiments of the present application, and are not intended to limit the present application in any form. Although the present application has been disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technical personnel in the field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present application. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present application without departing from the content of the technical solution of the present application are still within the scope of the technical solution of the present application.

Claims

1. An information processing method, It is characterized in that The method comprises: Obtaining object information, content information and background information, wherein the content information is information about the content to be recommended of the target object, and the background information is information about the scenario for recommending the content to be recommended; Inputting the object information, the content information and the background information into a target prediction model to predict the recommendation score of the target content, and outputting the predicted score of the target content under different tasks; Among them, the target prediction model is a sample target weight obtained by weight fusion based on sample personalized weights corresponding to different tasks and sample task weights, and is obtained by multi-task learning of a preset prediction network. The sample personalized weight corresponding to each task is generated by a personalized sub-network corresponding to each task based on the personalized characteristics of the object group, and the sample personalized weight corrects the sample task weight.

2. The method according to claim 1, It is characterized in that The target prediction model is trained by the following steps: Acquire a training sample, wherein the training sample includes sample object information, sample content information, sample background information and a scoring label; Inputting the sample object information, the sample content information and the sample background information into the prediction network for feature extraction to obtain corresponding sample object features, sample content features and sample background features; By using the prediction network, score prediction is performed based on the sample object features, the sample content features and the sample background features to obtain training scores of the training samples corresponding to different tasks; Perform a cross entropy operation based on the training score and the score label to determine a target loss corresponding to the prediction network; Based on the target loss, the prediction network is iteratively trained for multi-task learning of different tasks until the prediction network meets preset conditions to obtain a target prediction model.

3. The method according to claim 2, It is characterized in that The step of performing score prediction based on the sample object feature, the sample content feature and the sample background feature through the prediction network to obtain training scores of the training samples corresponding to different tasks includes: Through the prediction network, vector concatenation is performed on the sample object feature, the sample content feature and the sample background feature to obtain a sample common feature; The sample common features are respectively transmitted to the shared subnetwork and the task subnetwork corresponding to each task for feature crossover to obtain the sample shared features corresponding to the shared subnetwork and the sample task features corresponding to each task subnetwork; Perform feature concatenation based on the sample shared features and each of the sample task features to obtain a sample selection matrix; Based on the task gating and personalized sub-network corresponding to each task, weight calculation is performed to obtain the sample target weight corresponding to each task; Rating prediction is performed based on the sample target weight corresponding to each task and the sample selection matrix to obtain training scores of the training samples corresponding to different tasks.

4. The method according to claim 3, It is characterized in that The task gating and personalized sub-network corresponding to each task performs weight calculation to obtain the sample target weight corresponding to each task, including: The sample common features are transmitted to the task gate corresponding to each task for weighted calculation to obtain the sample task weight corresponding to each task; Generate a sample personalized weight corresponding to each task based on the personalized features of the object group through the personalized sub-network corresponding to each task, wherein the personalized features at least include the activity features of the object group; Based on the sample task weight and the sample personalized weight corresponding to each task, a bitwise multiplication operation is performed to obtain the sample target weight corresponding to each task.

5. The method according to any one of claims 2 to 4, It is characterized in that The performing a cross entropy operation according to the training score and the score label to determine a target loss corresponding to the prediction network includes: Perform a cross entropy calculation based on the training score and the score label corresponding to each of the tasks to determine the cross entropy loss corresponding to each task; Determine the task loss corresponding to each task based on the spatial parameters of the training samples and the cross entropy loss; Based on the loss weight corresponding to each of the tasks and the task loss, a target loss corresponding to the prediction network is determined.

6. The method according to claim 1, It is characterized in that The step of inputting the object information, the content information and the background information into a target prediction model to predict the recommendation score of the target content, and outputting the predicted score of the target content under different tasks, comprises: Inputting the object information, the content information and the background information into a target prediction model to extract features and obtain common features; The common features are respectively transmitted to the shared subnetwork and the task subnetwork corresponding to each task for feature crossover to obtain the shared features corresponding to the shared subnetwork and the task features corresponding to each task subnetwork; Generate a task weight corresponding to each task based on the common features through the task gating corresponding to each task; Generate a personalized weight corresponding to each task based on the personalized features of the object group through the personalized sub-network corresponding to each task; The score prediction is performed based on the task weight and personalized weight corresponding to each task to obtain the predicted score of each candidate content under different tasks.

7. The method according to claim 6, It is characterized in that The personalized sub-network may include three fully connected layers: a first fully connected layer, a second fully connected layer, and a third fully connected layer; the personalized weight corresponding to each task is generated based on the personalized features of the object group through the personalized sub-network corresponding to each task, including: Inputting the personalized features of the object group into the first fully connected layer of the personalized sub-network to obtain a first output; Inputting the first output into a second fully connected layer of a personalization sub-network to obtain a second output; The second output is input into the third fully connected layer of the personalized subnetwork to obtain the sample personalized weight.

8. The method according to claim 6, It is characterized in that The tasks include a first task and a second task, and the score prediction is performed according to the task weight and the personalized weight corresponding to each task to obtain the predicted score of the target content under each task, including: Determining a first target weight according to a task weight and a personalized weight corresponding to the first task; Determining a second target weight according to the task weight and the personalized weight corresponding to the second task; Performing score prediction based on the first target weight and the selection matrix corresponding to the first task to obtain a first predicted score for each candidate content under the first task; A score prediction is performed based on the second target weight and the selection matrix corresponding to the second task to obtain a second predicted score for each candidate content under the second task.

9. The method according to claim 7, It is characterized in that The method further comprises: determining a combined score based on the first prediction score and the second prediction score; Sort each candidate content in descending order according to the comprehensive score to obtain a content sequence; From the content sequence, candidate contents corresponding to a preset number of preceding comprehensive scores are determined as target contents for recommendation.

10. An information processing device, It is characterized in that The device comprises: An information acquisition module, used to acquire object information, content information and background information, wherein the content information is information about the content to be recommended of the target object, and the background information is information about the scenario for recommending the content to be recommended; A rating prediction model, used to input the object information, the content information and the background information into a target prediction model to predict the recommendation rating of the target content, and output the predicted rating of the target content under different tasks; Among them, the target prediction model is a sample target weight obtained by weight fusion based on sample personalized weights corresponding to different tasks and sample task weights, and is obtained by multi-task learning of a preset prediction network. The sample personalized weight corresponding to each task is generated by a personalized sub-network corresponding to each task based on the personalized characteristics of the object group.

11. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, which can be called by a processor to execute the method according to any one of claims 1 to 9.

12. A computer device, It is characterized in that include: Memory; One or more processors coupled to the memory; One or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and the one or more computer programs are configured to execute the method according to any one of claims 1 to 9.

13. A computer program product, It is characterized in that The computer program product includes a computer program, which is stored in a storage medium; a processor of a computer device reads the computer program from the storage medium, and the processor executes the computer program, so that the computer device executes the method according to any one of claims 1 to 9.