CN116383497A

CN116383497A - Material recommendation method and device with multiple recommendation targets and computer readable storage medium

Info

Publication number: CN116383497A
Application number: CN202310343876.5A
Authority: CN
Inventors: 刘松; 李亚辉; 高家华
Original assignee: Weimeng Chuangke Network Technology China Co Ltd
Current assignee: Weimeng Chuangke Network Technology China Co Ltd
Priority date: 2023-03-31
Filing date: 2023-03-31
Publication date: 2023-07-04

Abstract

The application provides a material recommendation method, device and equipment with multiple recommendation targets and a computer readable storage medium; the recommendation method of the multi-recommendation target comprises the following steps: acquiring first user characteristics of a target user and first material characteristics of candidate materials; according to the first user characteristics and the first material characteristics, respectively determining a predicted value of each recommended target in a plurality of recommended targets of the candidate materials by a target user; inputting the first user characteristic, the first material characteristic and the predicted value of each recommended target into a fusion network to carry out nonlinear weighted fusion of a plurality of predicted values, and obtaining fusion predicted values of a plurality of recommended targets. According to the method and the device, fusion of predicted values of personalized multi-recommendation targets is achieved, so that the probability that materials recommended based on the fused predicted values meet target user interest points is guaranteed to be higher, and recommendation accuracy is improved.

Description

Material recommendation method and device with multiple recommendation targets and computer readable storage medium

Technical Field

The present disclosure relates to the field of information flow recommendation technologies, and in particular, to a method and apparatus for recommending multiple recommendation targets, and a computer readable storage medium.

Background

In the information flow recommendation system, a user has various operation behaviors on materials. The recommendation system needs to predict the possibility of various operation behaviors of the user, comprehensively sorts the materials, improves the user experience quality and improves the accuracy of material throwing of the recommendation system. In the existing multi-recommendation-target recommendation system, materials to be recommended are generally determined in a fusion manner in a fixed manner based on predicted values of a plurality of recommendation targets. The influence of the user characteristics of different target users and the material characteristics of the materials to be recommended on the fusion of the predicted values of a plurality of recommended targets is not considered, and personalized recommendation can not be realized for the target users. The user may not be interested in pushing content and the accuracy of the recommendation is poor.

Disclosure of Invention

The application provides a material recommending method and device based on multiple recommending targets and a computer readable storage medium, which can realize personalized multiple-target recommending aiming at target users and improve the accuracy of recommending materials by a recommending system.

The technical scheme of the application is realized as follows:

in a first aspect, the present application provides a recommendation method for multiple recommendation targets, including: acquiring first user characteristics of a target user and first material characteristics of candidate materials; according to the first user characteristics and the first material characteristics, respectively determining a predicted value of each recommended target in a plurality of recommended targets of the candidate materials by the target user, wherein the predicted value of each recommended target is used for representing predicted information of user behaviors of the candidate materials by the target user, and the user behaviors correspond to each recommended target; inputting the first user characteristic, the first material characteristic and the predicted value of each recommended target into a fusion network to carry out nonlinear weighted fusion of a plurality of predicted values, so as to obtain fusion predicted values of a plurality of recommended targets; and the fusion predicted value is used for recommending candidate materials to the target user.

In some alternative embodiments, obtaining a first user characteristic of a target user and a first material characteristic of a candidate material includes: acquiring user information of a target user and material information of candidate materials; and inputting the user information and the material information into a plurality of shared networks to perform feature extraction to obtain a first user feature and a first material feature corresponding to each recommended target, wherein the plurality of shared networks are in one-to-one correspondence with the plurality of recommended targets.

In some alternative embodiments, determining a predicted value of the target user for each of a plurality of recommended targets of the candidate material according to the first user characteristic and the first material characteristic, respectively, includes: and inputting the first user characteristics and the first material characteristics corresponding to each recommended target into a plurality of independent tower networks to predict the recommended targets, and obtaining a predicted value of each recommended target, wherein the plurality of independent tower networks are in one-to-one correspondence with the plurality of recommended targets.

In some alternative embodiments, obtaining a first user characteristic of a target user and a first material characteristic of a candidate material includes: acquiring user information of a target user and material information of candidate materials; inputting the user information and the material information into a plurality of shared networks to perform feature extraction to obtain a second user feature and a second material feature corresponding to each recommended target, wherein the plurality of shared networks are in one-to-one correspondence with the plurality of recommended targets; and inputting the second user characteristics and the second material characteristics corresponding to each recommended target into a first sub-network in a plurality of independent tower networks for characteristic extraction to obtain the first user characteristics and the first material characteristics corresponding to each recommended target, wherein the plurality of independent tower networks are in one-to-one correspondence with the plurality of recommended targets.

In some alternative embodiments, determining a predicted value of the target user for each of a plurality of recommended targets of the candidate material according to the first user characteristic and the first material characteristic, respectively, includes: and inputting the first user characteristics and the first material characteristics output by the first sub-network of each independent tower network into a corresponding second sub-network to predict the recommended targets, and obtaining the predicted value of each recommended target.

In some alternative embodiments, after determining the predicted value of the target user for each of the plurality of recommended targets of the candidate material based on the first user characteristic and the first material characteristic, respectively, the method further comprises: acquiring behavior data of a target user on candidate materials; according to the behavior data, respectively determining a first deviation value of a predicted value of each recommended target; based on the first deviation value of the predicted value of each recommended target, parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target are updated.

In some alternative embodiments, updating parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target based on the first deviation value of the predicted value of each recommended target includes: based on the weight of the first deviation value, carrying out weighted summation on the first deviation value to obtain a second deviation value; calculating a first gradient value based on the second deviation value; parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target are updated based on the first gradient values.

In some optional embodiments, the method further includes, after inputting the first user characteristic, the first material characteristic, and the predicted value of each recommended target into the fusion network to perform nonlinear weighted fusion of a plurality of predicted values, obtaining the fused predicted value: determining a third deviation value of the fusion predicted value through a preset loss function; calculating a second gradient value based on the third deviation value; based on the second gradient value, parameters of the fusion network are updated.

In some alternative embodiments, the loss function is as shown in the following expression (1):

wherein loss () is the third bias value, prediction is the fusion predicted value, y _i Is the attribute value, mu, of the corresponding behavior of the ith recommended target _yi Is the attribute of the corresponding behavior of the ith recommended targetMean value of the values sigma _yi Is the variance of the attribute values of the corresponding behavior of the ith recommendation target.

In a second aspect, the present application provides a material recommendation device based on multiple recommendation targets, including: the first acquisition module is used for acquiring first user characteristics of a target user and first material characteristics of materials to be recommended; the prediction module is used for respectively determining the predicted value of each recommended target in a plurality of recommended targets of the candidate materials by the target user according to the first user characteristics and the first material characteristics, wherein the predicted value of each recommended target is used for representing the predicted information of the user behavior of the target user on the candidate materials, and the user behavior corresponds to each recommended target; the fusion module is used for inputting the first user characteristic, the first material characteristic and the predicted value of each recommended target into the fusion network to carry out nonlinear weighted fusion of a plurality of predicted values, so as to obtain fusion predicted values of a plurality of recommended targets; and the fusion predicted value is used for recommending candidate materials to the target user.

In some optional embodiments, the first obtaining module is further configured to obtain user information of the target user and material information of the candidate material; and inputting the user information and the material information into a plurality of shared networks to perform feature extraction to obtain a first user feature and a first material feature corresponding to each recommended target, wherein the plurality of shared networks are in one-to-one correspondence with the plurality of recommended targets.

In some optional embodiments, the prediction module is further configured to input the first user characteristic and the first material characteristic corresponding to each recommended target into a plurality of independent tower networks to predict the recommended target, so as to obtain a predicted value of each recommended target, where the plurality of independent tower networks are in one-to-one correspondence with the plurality of recommended targets.

In some optional embodiments, the first obtaining module is further configured to obtain user information of the target user and material information of the candidate material; inputting the user information and the material information into a plurality of shared networks to perform feature extraction to obtain a second user feature and a second material feature corresponding to each recommended target, wherein the plurality of shared networks are in one-to-one correspondence with the plurality of recommended targets; and inputting the second user characteristics and the second material characteristics corresponding to each recommended target into a first sub-network in a plurality of independent tower networks for characteristic extraction to obtain the first user characteristics and the first material characteristics corresponding to each recommended target, wherein the plurality of independent tower networks are in one-to-one correspondence with the plurality of recommended targets.

In some optional embodiments, the prediction module is further configured to input the first user characteristic and the first material characteristic output by the first sub-network of each independent tower network into a corresponding second sub-network to predict the recommended target, so as to obtain a predicted value of each recommended target.

In some optional embodiments, the apparatus further includes a first updating module, configured to obtain behavior data of the target user on the material to be recommended; according to the behavior data, respectively determining a first deviation value of a predicted value of each recommended target; based on the first deviation value of the predicted value of each recommended target, parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target are updated.

In some optional embodiments, the updating module is further configured to weight and sum the first deviation values based on weights of the first deviation values to obtain second deviation values; calculating a first gradient value based on the second deviation value; parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target are updated based on the first gradient values.

In some optional embodiments, the apparatus further comprises a second updating module, configured to determine a third deviation value of the fusion predicted value through a preset loss function; calculating a second gradient value based on the third deviation value; based on the second gradient value, parameters of the fusion network are updated.

In a third aspect, the present application provides a recommendation device with multiple recommendation targets, including: a memory for storing executable instructions or computer programs; and a processor configured to implement the method as provided in the first aspect of the present application when executing the executable instructions stored in the memory.

In a fourth aspect, the present application provides a computer readable storage medium storing executable instructions or a computer program for implementing a method as provided in the first aspect of the present application when executed by a processor.

Compared with the prior art, the technical scheme provided by the application has the beneficial effects that:

in the application, nonlinear weighted fusion of a plurality of predicted values is performed through a fusion network based on first user characteristics of a target user, first material characteristics of materials to be recommended and predicted values of a plurality of recommended targets. In this way, in the process of fusing the predicted values of the plurality of recommended targets through the fusion network, the first user characteristics of the target users are considered, the first material characteristics of the candidate materials are considered, the fusion network combines the user information, the material information and the predicted information of the behavior of the user on the materials, and on the basis of the granularity of the user, the predicted values of the plurality of recommended targets are unified into one fused predicted value in a nonlinear weighting mode. Aiming at each target user, based on the difference of the first user characteristics and the first material characteristics, the nonlinear weighting result of the fusion network is different, which is equivalent to that each target user has a set of weighting modes belonging to a plurality of recommended targets of the target user, so that the fusion of the predicted values of the personalized multi-recommended targets is realized, the probability that the materials recommended based on the fused predicted values meet the interest points of the target user is higher, and the recommendation accuracy is improved.

Drawings

FIG. 1 is a flow chart of a recommendation method of multiple recommendation targets in the related art;

FIG. 2 is a schematic architecture diagram of a multi-recommendation-target recommendation system according to an embodiment of the present application;

fig. 3 is a schematic structural diagram of a server according to an embodiment of the present application;

FIG. 4 is a schematic flow chart of a material recommendation method based on multiple recommendation targets according to an embodiment of the present application;

fig. 5 is a schematic flow chart of a material recommendation method based on multiple recommendation targets applied to content sharing application software according to an embodiment of the present application;

FIG. 6 is a schematic structural diagram of a recommendation model according to an embodiment of the present application;

fig. 7 is a schematic structural diagram of a server according to an embodiment of the present application.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present application more apparent, the present application will be described in further detail with reference to the accompanying drawings, and the described embodiments should not be construed as limiting the present application, and all other embodiments obtained by those skilled in the art without making any inventive effort are within the scope of the present application.

In the following description, reference is made to "some embodiments" which describe a subset of all alternative embodiments, but it is to be understood that "some embodiments" can be the same subset or different subsets of all alternative embodiments and can be combined with one another without conflict.

In the following description, the terms "first/second/third" are used merely to distinguish between similar objects and do not represent a particular ordering of the objects, it being understood that the "first/second/third" may be interchanged with a particular order or precedence where allowed, to enable embodiments of the present application described herein to be implemented in other than those illustrated or described herein.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of the present application only and is not intended to be limiting of the present application.

In order to illustrate the technical solutions described in the present application, the following description is made by specific examples.

With rapid development of technology, especially wide spread of mobile communication networks and mobile terminals, various material contents exist on the networks. Users can also produce and make material content anytime and anywhere, and different service manufacturers provide platforms. The service side collects a part of material content produced by the user on one hand, and screens high-quality material content in different modes on the other hand. And pushing corresponding material content to the user based on different user interest types so as to improve the probability of clicking and viewing the pushed material content data by the user.

On each network platform, a large amount of material is typically stored, forming a material database. The content data of these materials may include multimedia data in various formats. For example, video data, audio data, image data, text data, etc. The form of which varies in different business scenarios. Such as live programming, short videos, songs, comics, audio novels, news, articles, food-item platforms, and the like. The service platform side can actively put the multimedia data on the platform, for example, the platform side can purchase various multimedia data from the copyright side and display the multimedia data on the platform. Alternatively, registered users on the platform can also produce various types of multimedia data through the electronic device and upload the produced multimedia data to the platform, and other users can also search and browse the multimedia data through the electronic device. The electronic device includes, but is not limited to, various types of user intelligent terminals such as notebook computers, tablet computers, desktop computers, mobile terminals and the like.

There may be a plurality of different forms of multimedia data for different traffic scenarios. For example, in a live broadcast service scene, a user can use a camera to collect video data, and perform operations such as beautifying and wheat connection on the video data so as to generate a live broadcast program; in the short video service scene, a user can acquire video data by using a camera, and operations such as beautifying, clipping, adding special effects and the like are performed on the video data, so that a short video is generated, and in the information sharing communication service scene, the user can upload edited characters, images, videos and the like to generate blog information and the like. The multimedia data can be published on the corresponding platform, other users can use intelligent equipment such as terminals and the like to acquire the data on the corresponding platform, or operations such as sharing the data content can be performed.

The information flow recommending system is used for more accurately screening out contents possibly interested by a user from massive data and recommending the contents to the user. The recommendation system can achieve the purposes of pulling new and pulling the moon through the materials interested by the user, and further achieve the growth of the user. The method is essentially used for recommending the materials with high click rate to the user, and is widely applied to the fields of electronic commerce, searching, advertising and the like and used for recommending personalized materials for the user. For example, in an advertising scenario, a personalized recommendation system may push material with advertisements to a user through the user's features and preferences, etc. If the user finally generates click conversion behavior, the advertisement pushing result can be considered, otherwise, the pushing fails.

In the recommendation system, a plurality of operation behaviors exist for materials by a user. For example, the user clicks, browses, interacts with the recommended materials, and the plurality of operation behaviors can correspond to a plurality of recommended targets. The recommendation system needs to predict the possibility that a user implements various behaviors on the materials to be recommended, and sort the materials to be recommended based on the possibility, so as to recommend the materials to the user. Therefore, the accuracy of recommending materials by the recommending system is improved, and user experience is further improved.

Currently, in a recommendation system, a fixed weight is generally assigned to a plurality of recommendation targets based on experience, so that a plurality of recommendation target predicted values are weighted, and ranking of candidate materials is achieved. Referring to fig. 1, fig. 1 is a flowchart of a recommendation method of multiple recommendation targets in the related art. Firstly, a recommendation system acquires user information of a target user and material information of candidate materials (S1); extracting characteristics of the user information and the material information through a sharing network to obtain user characteristics and material characteristics (S2); then, predicted values of the plurality of recommended targets are output through the plurality of independent tower networks corresponding to the plurality of recommended targets, respectively (S3).

However, for multi-recommendation target recommendation systems, each target user is differentiated for the behavior of different recommended materials at different time periods in different scenes. And the recommended targets corresponding to different behaviors have different weights for different users. Therefore, a fixed weight is allocated to a plurality of recommendation targets, and materials to be recommended, which are determined from candidate materials, are weighted, so that a user may not be interested, and the accuracy of recommendation is poor.

In order to solve the above problems, embodiments of the present application provide a recommendation method, apparatus and computer readable storage medium for multiple recommendation targets, which can implement personalized recommendation for each target user, and improve accuracy of recommending materials by a recommendation system.

The recommendation method of multiple recommendation targets provided by the embodiment of the application can be implemented by various electronic devices, for example, can be implemented by a server alone or can be implemented by a terminal and the server cooperatively. For example, the server alone performs a recommendation method of multiple recommendation targets described below, or the terminal transmits a recommendation request message to the server, and the server performs a recommendation method of multiple recommendation targets according to the received recommendation request message.

The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, basic cloud computing services such as big data and an artificial intelligent platform. The terminal may be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc. The terminal and the server may be directly or indirectly connected through wired or wireless communication, which is not specifically limited in the embodiment of the present application.

Referring to fig. 2, fig. 2 is a schematic architecture diagram of a multi-recommendation system 200 provided in an embodiment of the present application, where terminals (a terminal 201-1 and a terminal 201-2 are exemplarily shown) are connected to a server 203 through a network 202, and the network 202 may be a wide area network or a local area network, or a combination of the two.

In some embodiments, the terminal or the server 200 may implement the multi-recommendation target recommendation method provided in the embodiments of the present application by running a computer program. For example, the computer program may be a native program or a software module in an operating system; for example, a local (APP) application program (APP) may be used, that is, a program that needs to be installed in an operating system to be run, such as a live APP, an APP of a content sharing class, and the like; the method can also be an applet, namely a program which can be run only by being downloaded into a browser environment; but also an applet that can be embedded in any APP. In general, the computer program may be any form of application program, module, or plug-in, which is not specifically limited by the embodiments of the present application.

Here, the terminal 201-1 is described as a target user, and the terminal 201-1 displays an interface of the APP on the current interface 210-1. For example, terminal 201-1 may display the recommended materials presented by the server to the target user on the current interface. The server 200 may predict a plurality of recommended targets according to the obtained user characteristics of the target user and the material characteristics of the material to be recommended, so as to obtain predicted values of the plurality of recommended targets. And inputting the user characteristics, the material characteristics and the obtained predicted values of the plurality of recommended targets into a fusion network, fusing the predicted values of the plurality of recommended targets into a value, and selecting materials from the materials to be recommended to recommend to the target user through the fused predicted values.

Referring to fig. 3, fig. 3 is a schematic structural diagram of a server provided in an embodiment of the present application, and the server 203 shown in fig. 3 may include at least one processor 310, a memory 320, at least one network interface 330, and a user interface 340. The various components in device 300 are coupled together by bus system 350. It is understood that bus system 350 is used to enable connected communications between these components. The bus system 350 includes a power bus, a control bus, and a status signal bus in addition to the data bus. But for clarity of illustration the various buses are labeled in fig. 3 as bus system 350.

The processor 310 may be an integrated circuit chip with signal processing capabilities such as a general purpose processor, which may be a microprocessor or any conventional processor, or the like, a digital signal processor (Digital Signal Processor, DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, or the like.

The user interface 340 includes one or more output devices 341 that enable presentation of media content, including one or more speakers and/or one or more visual displays. The user interface 304 also includes one or more input devices 342, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

The memory 320 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid state memory, hard drives, optical drives, and the like. Memory 320 optionally includes one or more storage devices physically remote from processor 310.

Memory 320 includes volatile memory or nonvolatile memory, and may also include both volatile and nonvolatile memory. The non-volatile Memory may be a Read Only Memory (R OM), and the volatile Memory may be a random access Memory (Random Access Memory, RAM). The memory 320 described in embodiments of the present application is intended to comprise any suitable type of memory.

In some embodiments, memory 320 is capable of storing data to support various operations, examples of which include programs, modules and data structures, or subsets or supersets thereof, as exemplified below.

An operating system 321 including system programs for handling various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and handling hardware-based tasks;

network communication module 322 for reaching other computing devices via one or more (wired or wireless) network interfaces 330, the exemplary network interfaces 330 comprising: bluetooth, wireless compatibility authentication (WiFi), and universal serial bus (Universal Serial Bus, USB), etc.;

A presentation module 323 for enabling presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 342 (e.g., a display screen, speakers, etc.) associated with the user interface 340;

an input processing module 324 for detecting one or more user inputs or interactions from one of the one or more input devices 342 and translating the detected inputs or interactions.

In some embodiments, the multi-recommendation recommending apparatus provided in the embodiments of the present application may be implemented in software, and fig. 3 shows a multi-recommendation recommending apparatus 325 stored in a memory 320, which may be software in the form of a program and a plug-in, and includes the following software modules: the first acquisition module 3251, the prediction module 3252, and the fusion module 3253 are logical, and thus may be arbitrarily combined or further split according to the implemented functions. The functions of the respective modules will be described hereinafter.

In other embodiments, the multi-recommendation recommending apparatus provided in the embodiments of the present application may be implemented in hardware, and by way of example, the multi-recommendation recommending apparatus provided in the embodiments of the present application may be a processor in the form of a hardware decoding processor that is programmed to perform the multi-recommendation recommending method provided in the embodiments of the present application, for example, the processor in the form of a hardware decoding processor may employ one or more application specific integrated circuits (Application Specific Integrated Circuit, ASIC), DSP, programmable logic device (Programmable Logic Device, PLD), complex programmable logic device (C omplex Programmable Logic Device, CPLD), field programmable gate array (Field-Progr ammable Gate Array, FPGA), or other electronic components.

The following describes a multi-recommendation-object recommendation method provided in the embodiment of the present application in conjunction with exemplary applications and implementations of the multi-recommendation-object recommendation device provided in the embodiment of the present application.

Referring to fig. 4, fig. 4 is a schematic flow chart of a material recommending method based on multiple recommending targets according to an embodiment of the present application, and the following description will refer to the steps shown in fig. 4.

S401, acquiring first user characteristics of a target user and first material characteristics of candidate materials;

wherein the first user characteristic may comprise a user attribute characteristic of the target user. By way of example, the user attributes of the target user may include the age, gender, ethnicity, race, etc. of the target user, and may also include the target user's position, constellation, hobbies of interests, religious beliefs, etc.

In some embodiments, in S401, after obtaining the user attribute of the target user, the recommendation system may perform feature extraction on the user attribute to obtain the first user feature. For example, the data describing the user attributes may be a long text. The recommender system may perform feature extraction on the segment of text to obtain a first user characteristic of the target user.

In some embodiments, the user attribute may not be limited to text information, but may be data of the type of image, video, or the like. For example, containing pictures and videos of the target user, the server may obtain skin tone, age, gender, etc. of the target user based on image recognition and analysis. User attribute information such as ethnicity of a target user is acquired based on dressing of the user in the image and video.

In some embodiments, the server may acquire the user attribute information mentioned above, or the server may further analyze information determined by acquiring interest preferences of the target user based on usage habits of the target user. For example, for a content sharing platform, a target user needs to register when using the platform, and the user attribute information may be information of user registration authentication provided by the target user. Or the server can also acquire the content data which is actively inquired or frequently clicked and checked by the target user, and can analyze the possible interest and hobbies of the target user according to the type of the content data. For example, the target user is often concerned with the content of sports games, such as ball games and the like. Then the target user can be analyzed for a high probability of having a preference for this type of sports. Or the server may also obtain the user attribute information based on any other channel. Regarding the channel from which such data is obtained, the embodiments of the present application are not particularly limited thereto.

Based on the same method for acquiring the first user characteristics, the server acquires the first material characteristics of the candidate materials, or extracts the content information of the candidate materials to acquire the characteristic information capable of representing the content of the candidate materials.

Illustratively, taking the content sharing platform as an example, the candidate material may be text content, image content, or video content. Or may be one or more of text content, image content and video content. The acquiring of the first material characteristic may be performing text and/or image content recognition on the candidate material, so as to acquire information such as one or more keywords capable of representing the candidate content. For example, the candidate material is video data. The first material characteristics may include, but are not limited to, characteristics of style, type, duration of video, etc.

S402, respectively determining a predicted value of each recommended target in a plurality of recommended targets of the candidate materials by a target user according to the first user characteristics and the first material characteristics; the predicted value of each recommended target is used for representing predicted information of the target user on the user behavior of the candidate material, and the user behavior corresponds to each recommended target.

The plurality of recommended targets may be a click rate, an interaction rate, a consumption time period, etc. of the user on the material.

In some embodiments, multiple recommendation targets for the target user may be determined based on the target user's behavior on the historical recommended materials. The behavior of the target user on the historical recommended material may include the behavior of the target user on the historical recommended material, such as praise, collection, forwarding, comment, and the like.

Illustratively, a content sharing platform is taken as an example. The target user logs in the platform through the terminal, the server selects materials to be recommended from the candidate materials as materials to be recommended and sends the materials to the corresponding terminal, and the terminal displays the received materials through the display interface. The materials received by the terminal can be any type of materials such as characters, pictures, videos and the like. The operation behavior of the target user on the history recommended material may include: clicking one or more materials to view, and performing operations such as praying, collection, forwarding, comment and the like on the content of one or more materials.

It can be appreciated that the operational behavior of the target user on the historically recommended materials can reflect to some extent the degree of interest of the target user on the historically recommended materials. While different target users may have different operational behaviors for the same content of interest. Therefore, for the recommendation system, selecting a material to be recommended from candidate materials to be recommended to a target user requires prediction of the possibility of various operation behaviors of the target user. That is, determining the predicted value of each of the plurality of recommended targets corresponding to the candidate material predicts the probability that the target user may have an operation behavior corresponding to the recommended target after recommending the candidate material to the target user. For example, aiming at the click action of the target user on the recommended material, the predicted value is the probability of the recommended target of the click action of the target user. Or aiming at the forwarding behavior of the target user on the recommended materials, wherein the predicted value is the probability of the recommended target of the forwarding behavior of the target user.

In addition, after the candidate materials are recommended to the target users, the operation behaviors of each target user on the recommended materials have correlation with the attribute characteristics of the target users and the material characteristics of the recommended materials. That is, in the step S402, the process of determining the predicted value of each of the plurality of recommended targets corresponding to the candidate material according to the first user characteristic and the first material characteristic is to analyze what first user characteristic the corresponding target user has and what first material characteristic the corresponding historical recommended material has according to the operation behavior of the target user on the historical recommended material. And determining the predicted value of the corresponding recommended target based on the first user characteristic of the current target user and the first material characteristic of the candidate material.

Taking 500 target users and 1000 historical recommended materials as examples, the 500 target users can respectively acquire user characteristics with multiple dimensions, such as gender, age, ethnicity and the like. Likewise, for 1000 historical recommended materials, the material characteristics of multiple dimensions can be obtained respectively. Such as the type of material, style, etc. Taking as an example that each target user has N user characteristics and each history recommended material has M material characteristics. The user characteristics of 500 target users and the material characteristics of 1000 historical recommended materials are respectively expressed in the form of arrays, so that two arrays of 500 XN and 1000 XM can be obtained. For each operational behaviour, the user profile and the corresponding material profile have 500000×n×m possible outcomes, respectively. That is, the predicted value of each recommended target is determined by determining the ratio of the first user characteristic of the corresponding target user and the first material characteristic of the candidate material to the total number of results from the result of 500000×n×m.

In some embodiments, for one of the plurality of recommended targets, the target user may not have an operational behavior corresponding to the corresponding recommended target. Therefore, in order to reduce the calculation amount in the process of determining the predicted value of each recommended target, the corresponding first user characteristic and the first material characteristic can be acquired based on each recommended target, so as to reduce the dimensions of the first user characteristic and the first material characteristic.

Based on this, in some embodiments, the multi-recommendation target recommendation method may further include: acquiring user information of a target user and material information of candidate materials; and inputting the user information and the material information into a plurality of shared networks to perform feature extraction, and obtaining a first user feature and a first material feature corresponding to each recommended target.

In some embodiments, the plurality of shared networks may be network models for extracting key information from various information. Such as deep neural networks, fully connected neural networks, etc. Based on the difference of each recommended target, the plurality of shared networks can acquire different first user characteristics and first material characteristics aiming at target users and candidate materials respectively.

For example, again taking the 500 target user samples and 1000 historical recommended materials samples described above as examples, the user features and corresponding materials features each have 500000×n×m possible outcomes. For a specific recommended target, if 250 target users do not have operation behaviors corresponding to the recommended target, after the target users are screened through the shared network, the total result number required to be calculated can be reduced by half, and the calculation amount can be greatly reduced.

In some embodiments, the plurality of shared networks are in one-to-one correspondence with a plurality of recommended targets. The plurality of shared networks may be network models that are obtained based on historical behavior data of the target user as sample data. Or may be a general network trained based on historical behavioral data of a plurality of users. For example, taking a content sharing platform as an example, the server may train to obtain a corresponding shared network based on all users in the platform, or may select a part of users from all users, and use historical behavior data of a plurality of selected users as samples. Or, the historical behavior data of a plurality of users and the target user can be taken as samples at the same time, and the corresponding shared network can be obtained through training. For example, the server may train to obtain a general network based on historical data of a plurality of platform users, and then, for each target user, adjust the general network through historical behaviors of the target user to obtain a personalized shared network corresponding to each target user.

In the embodiment of the invention, the first user characteristic and the first material characteristic which have relevance with each recommended target can be respectively obtained through the shared network corresponding to each recommended target in the plurality of recommended targets. And then, according to the first user characteristics and the first material characteristics, determining the predicted value of each recommended target, so that calculation irrelevant to the corresponding recommended target can be reduced, and the calculation efficiency of the server is greatly improved.

In some embodiments, the first user characteristic and the first material characteristic associated with the plurality of recommendation targets are obtained through the plurality of sharing networks. The predicted value of each recommended target may be obtained based on the same method as described above for determining the predicted value of each recommended target separately. Alternatively, the server may obtain the predicted value of each recommended target through a preset network model.

In some embodiments, the above method may further comprise: and inputting the first user characteristics and the first material characteristics corresponding to each recommended target into a plurality of independent tower networks to predict the recommended targets, and obtaining a predicted value of each recommended target, wherein the plurality of independent tower networks are in one-to-one correspondence with the plurality of recommended targets.

In some embodiments, the independent tower network may be a network model trained using training samples. Such as deep neural networks, fully connected neural networks, etc. The training samples may be historical behavior data of the target user or may be based on historical behavior data of multiple users. For example, using a content sharing platform as an example, the server may be based on all users within the platform, or may select a portion of the users from all users. And training and obtaining corresponding independent tower networks by taking the historical behavior data of the selected multiple users as samples. Or the historical behavior data of a plurality of users and the target user can be taken as samples at the same time, and the corresponding independent tower network is obtained through training. For example, the server may train to obtain a general network based on historical data of a plurality of users, and then, for each target user, adjust the general network through historical behaviors of the target user to obtain a personalized independent tower network corresponding to each target user.

In some embodiments, for one of the plurality of recommended targets, a predicted value for each recommended target is obtained, and each recommended target is different in emphasis on the user characteristics and the material characteristics, except that the target user may not have an operational behavior corresponding to the recommended target. For example, the candidate material is video content and the recommendation target is the length of time the user consumes (i.e., the length of time the target user views the video). The duration of the video content may not be correlated to the gender characteristics of the target user. Thus, in determining the predicted value of the recommended target for the consumption period, it may not be necessary to determine the gender characteristics of the target user. Or may be other first user characteristics or material characteristics that have no relevance to the recommended goals. Therefore, in order to reduce the amount of calculation in determining the predicted value of each recommended target as much as possible, the influence between recommended targets may be considered separately for each recommended target. The corresponding first user characteristics and the first material characteristics are obtained, so that the dimensions of the first user characteristics and the first material characteristics are reduced.

For example, the 500 target users and 1000 historical recommended materials described above are also taken as examples. After the target users without corresponding operation behaviors are excluded through the sharing network, the number of results of the user features and the corresponding material features is reduced from 500000×n×m to 250000×n×m. If the situation that the relevance between different features and the recommended target is low is considered, for example, only half of N user features and M material features respectively have relevance with the recommended target. After the user characteristics and the material characteristics are screened through the independent tower network, the number of results of the user characteristics and the corresponding material characteristics can be reduced by four times. The calculation amount of the server can be obviously reduced under the condition of not influencing the predicted value result.

In some embodiments, the multi-recommendation-target-based material recommendation method may further include: acquiring user information of a target user and material information of candidate materials; inputting the user information and the material information into a plurality of shared networks to perform feature extraction, and obtaining a second user feature and a second material feature corresponding to each recommended target; wherein, a plurality of sharing networks are in one-to-one correspondence with a plurality of recommendation targets; and inputting the second user characteristics and the second material characteristics corresponding to each recommended target into a first sub-network in an independent tower network corresponding to each recommended target for characteristic extraction to obtain the first user characteristics and the first material characteristics corresponding to each recommended target, wherein the independent tower networks are in one-to-one correspondence with the recommended targets.

In some embodiments, the first subnetwork may be a separate network model in a plurality of independent tower networks, or the first subnetwork may also be one or more layers of network structure in a corresponding independent tower network. For example, the first subnetwork may be a separate network model connected to the corresponding independent tower network. After the first sub-network obtains the second user features and the second material features output by the plurality of shared networks, the second user features and the second material features are subjected to feature extraction again based on the relevance between the second user features and the second material features and the corresponding recommended targets, so that the first user features and the first material features are obtained. Alternatively, the independent tower network may be a multi-layered neural network model. A fully connected neural network, for example, of three layers, includes an input layer, a hidden layer, and an output layer. The first subnetwork may be in a three-layer network structure, one or more of which.

In the embodiment of the invention, the first sub-network in the independent tower network corresponding to the plurality of recommended targets is utilized to respectively extract the second user characteristics and the second user characteristics acquired by the plurality of shared networks again, so that the calculation irrelevant to the corresponding recommended targets can be reduced, and the calculation efficiency of the server is greatly improved.

In some embodiments, after the first user characteristic and the first material characteristic associated with each recommended target are obtained through the first sub-network in each independent tower network, the predicted value of each recommended target may be obtained according to the same method for determining the predicted value of each recommended target respectively. Alternatively, the server may obtain the predicted value of each recommended target through a preset network model.

In some embodiments, the multi-recommendation-target-based material recommendation method may further include: and inputting the first user characteristics and the first material characteristics output by the first sub-network of each independent tower network into a corresponding second sub-network to predict the recommended targets, and obtaining the predicted value of each recommended target.

Wherein the second sub-network may be an independent network model in an independent tower network corresponding to each recommended target. Or the second subnetwork may also be one or more of the corresponding individual tower networks. For example, the second subnetwork may be a separate network model connected to a corresponding independent tower network, which is composed of the second subnetwork and the first subnetwork. Or the independent tower network may be a multi-layered neural network model. A fully connected neural network, for example, of three layers, includes an input layer, a hidden layer, and an output layer. The second subnetwork may be in a three-layer network structure, one or more of which.

In the embodiment of the application, nonlinear weighted fusion of a plurality of predicted values is performed through a fusion network based on the first user characteristics of the target user, the first material characteristics of the candidate materials and the predicted values of a plurality of recommended targets. In this way, in the process of fusing the predicted values of the plurality of recommended targets through the fusion network, the first user characteristics of the target users are considered, the first material characteristics of the candidate materials are considered, the fusion network combines the user information, the material information and the predicted information of the behavior of the user on the materials, and on the basis of the granularity of the user, the predicted values of the plurality of recommended targets are unified into one fused predicted value in a nonlinear weighting mode. Aiming at each target user, based on the difference of the first user characteristics and the first material characteristics, the nonlinear weighting result of the fusion network is different, which is equivalent to that each target user has a set of weighting modes belonging to a plurality of recommended targets of the target user, so that the fusion of the predicted values of the personalized multi-recommended targets is realized, the probability that the materials recommended based on the fused predicted values meet the interest points of the target user is higher, and the recommendation accuracy is improved.

In some embodiments, the shared network and the independent tower network are each a network model that is pre-trained based on sample data. For a target user, the operational behavior of the target user may vary over time. Therefore, if the fixed network model is used to obtain the corresponding user characteristics and material characteristics so as to obtain the predicted values of different recommended targets, the deviation value of the predicted values may gradually increase along with the change of the operation behaviors of the target users. Thus, for each target user, the parameters of the corresponding network model may be updated based on the actual behavior data of the target user to reduce the deviation of the predicted values.

In some embodiments, the multi-recommendation-target-based material recommendation method may further include: acquiring behavior data of a target user on candidate materials; according to the behavior data, respectively determining a first deviation value of a predicted value of each recommended target; based on the first deviation value of the predicted value of each recommended target, parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target are respectively updated.

In some embodiments, the server may determine to update parameters of the shared network or the independent tower network based on the actual calculation results, or update parameters of the shared network and the independent tower network simultaneously. For example, the prediction values of the plurality of recommended targets may have errors because, when feature extraction is performed through the shared network, the first user feature and the first material feature associated with the corresponding recommended target are extracted to have errors. At this time, the server may update the parameters of the shared network. Or the predicted values of the plurality of recommended targets may be error-prone, or may be error-prone when determining the predicted values of the plurality of recommended targets based on the independent tower network. At this point, the server may update the parameters of the independent tower network. Or the predicted values of the plurality of recommended targets have errors, and the errors of the predicted values of the plurality of recommended targets can be determined based on the independent tower network because the errors of the first user characteristics and the first material characteristics associated with the corresponding recommended targets are extracted when the characteristics are extracted through the shared network. At this time, the server may update parameters of the shared network and the independent tower network at the same time.

In the embodiment of the application, the first deviation value corresponding to the predicted value of each recommended target is determined through the behavior data of the target user on the candidate materials, and the parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target are updated through the first deviation value corresponding to the predicted value of each recommended target. Therefore, parameters of the corresponding shared network and/or independent tower network can be timely adjusted, so that the shared network and/or the independent tower network can be adapted to the corresponding target user, and deviation of the predicted value is reduced.

In some embodiments, the step of updating the parameters of the plurality of target networks, respectively, according to the first deviation value may include: a weight based on a first bias value of the predicted value of each recommended target; carrying out weighted summation on each first deviation value based on the weight of each first deviation value to obtain a second deviation value; calculating a first gradient value based on the second deviation value; parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target are updated based on the first gradient values.

The order of execution of the parameter updates of the shared network and the independent tower network may not be limited. In the embodiment of the application, the parameters of the shared network can be updated first, and then the parameters of the independent tower network can be updated. Or the independent tower network parameters can be updated first, and then the shared network parameters can be updated. Or parameters of the shared network and the independent tower network may also be updated simultaneously.

S403, inputting the first user characteristic, the first material characteristic and the predicted value of each recommended target into a fusion network to perform nonlinear fusion of a plurality of predicted values, so as to obtain fusion predicted values of a plurality of recommended targets; the fusion predicted value is used for recommending candidate materials to a target user.

It will be appreciated that for the same candidate material, the predicted values for a plurality of recommended targets may be obtained by the method described above. Meanwhile, for a plurality of candidate materials, the predicted values of a plurality of corresponding recommended targets can be obtained respectively. And fusing the predicted values of the plurality of recommended targets into one value, wherein the obtained fused predicted values can reflect the common probability of the predicted values of the plurality of recommended targets. That is, the probability of whether the recommendation to the target user is successful can be reflected, so that the recommendation method can be used for recommending candidate materials to the target user. In addition, when recommending the candidate materials to the target user, in order to determine the order of the candidate materials, the fused predicted values can also be used for sorting the candidate materials, so that the materials are sequentially recommended to the target user.

The fusion network can be obtained by training based on behavior information of the target user on the historical recommended materials, user characteristics of the target user and material characteristics of the historical recommended materials as sample data. The converged network may be an existing network model, such as a deep neural network model, or a fully connected neural network model, or the like.

In some embodiments, since the converged network is a network model that is pre-trained based on sample data. For a target user, the operational behavior of the target user may vary over time. Therefore, if the fused predicted values of the plurality of recommended targets are obtained using a fixed network model, the deviation value of the fused predicted values may gradually increase as the operation behavior of the target user changes. Therefore, for each target user, updating the parameter correspondence of the converged network is required to reduce the deviation of the converged predicted value.

In some embodiments, the parameter update procedure for the converged network may include the steps of: determining a third deviation value of the fused predicted value through a preset loss function; calculating a second gradient value based on the third deviation value; based on the second gradient value, parameters of the fusion network are updated.

Wherein the loss function can be expressed as the following expression (1):

/>

wherein loss () is the third bias value, prediction is the fusion predicted value, y _i Is the attribute value, mu, of the corresponding behavior of the ith recommended target _yi Is the mean value, sigma, of the attribute values of the corresponding behavior of the ith recommended target _yi Is the variance of the attribute values of the corresponding behavior of the ith recommendation target.

In this embodiment of the present application, the classification problem (for example, whether the target user clicks, whether the target user interacts) and the regression problem (for example, the consumption duration of the target user, etc.) are normalized by the loss function. The unified calculation is convenient, and the problem of inconsistent weights of different recommendation targets is solved. Personalized recommendation aiming at target users is realized, and the accuracy of recommendation is improved.

In the following, an exemplary application of the embodiments of the present application in a practical application scenario will be described.

When the material recommending method based on the multiple recommending targets is applied to content sharing application software, such as microblog, at this time, the target user can be any registered user in the microblog, and the candidate materials can include blogs, pictures, videos and the like released by a blogger. The implementation flow of the material recommendation method based on multiple recommendation targets applied to the content sharing application software can be referred to fig. 5, and fig. 5 is a schematic flow diagram of the application of the material recommendation method based on multiple recommendation targets applied to the content sharing application software. The steps shown in fig. 5 will be described below.

S501, a server acquires user information of a target user and material information of candidate materials;

s502, the server inputs user information and material information into a plurality of shared networks;

s503, the server obtains a second user characteristic and a second object characteristic corresponding to each recommended target;

s504, the server inputs the second user characteristics and the second object characteristics corresponding to each recommended target into a first sub-network in a plurality of independent tower networks to perform characteristic extraction;

S505, the server obtains first user characteristics and first material characteristics corresponding to each recommendation target;

s506, the first user characteristics output by the first sub-network of each independent tower network and the second sub-network corresponding to the first material characteristics are input to the server to conduct recommendation target prediction;

s507, the server obtains the predicted value of each recommended target;

s508, the server inputs the first user characteristics, the first material characteristics and the predicted value of each recommendation target into a fusion network for fusion;

s509, the server acquires the fusion predicted value.

The above-described multi-recommendation-object-based material recommendation method is described below with one recommendation model provided in a server. Fig. 6 is a schematic structural diagram of a recommendation model according to an embodiment of the present application. Referring to fig. 6, taking three recommended targets as an example, the recommendation model includes: a shared network (shared network 1, shared network 2, and shared network 3 shown in fig. 6) corresponding to the three recommended targets, three independent tower networks (independent tower network 1, independent tower network 2, and independent tower network 3 shown in fig. 6) corresponding to the three recommended targets, and one converged network, respectively. Wherein each independent tower network comprises a first subnetwork (first subnetwork 1, first subnetwork 2 and first subnetwork 3 shown in fig. 6) and a second subnetwork (second subnetwork 1, second subnetwork 2 and second subnetwork 3 shown in fig. 6), respectively.

In the model, after the server acquires the user information of the target user and the material information of the candidate materials, the user information and the material information are respectively input into three sharing networks. And extracting features through three sharing networks to respectively obtain a second user feature and a second material feature corresponding to each recommended target. The output of the three sharing networks (namely the second user characteristic and the second material characteristic) are respectively used as the input of each first sub-network, and the characteristic extraction is carried out again through the first sub-network, so that the first user characteristic and the first material characteristic corresponding to each recommended target are respectively obtained. And taking the output (namely the first user characteristic and the first material characteristic) of the first sub-network in each independent tower network as the input of a corresponding second sub-network, and carrying out recommendation target prediction through the second sub-network to respectively obtain the prediction values of three recommendation targets. The output of each first sub-network (namely, the first user characteristic and the first material characteristic) and the output of each second sub-network (namely, the predicted values of three recommended targets) are respectively input into a fusion network, and nonlinear weighted fusion is carried out on the predicted values of the recommended targets through the fusion sharing network to obtain fusion predicted values.

By means of the recommendation model, the multi-recommendation-object-based material recommendation method provided by the embodiment of the application can be achieved, for example, the multi-recommendation-object recommendation method shown in the above-mentioned fig. 4 or 5.

Continuing with the description below of an exemplary architecture implemented as a software module for the multi-recommendation device 325 provided by embodiments of the present application, in some embodiments, still as shown in fig. 3, the software modules stored in the data processing device 325 of the memory 320 may include: a first obtaining module 3251, configured to obtain a first user characteristic of a target user and a first material characteristic of a candidate material; the prediction module 3252 is configured to determine, according to the first user characteristic and the first material characteristic, a predicted value of each of a plurality of recommended targets of the candidate material by the target user, where the predicted value of the recommended target is used to characterize predicted information of a behavior of the user corresponding to the recommended target; the fusion module 3253 is configured to input the first user characteristic, the first material characteristic, and the predicted value of each recommended target into a fusion network to perform nonlinear weighted fusion of multiple predicted values, so as to obtain a fused predicted value of multiple recommended targets; and the fusion predicted value is used for recommending candidate materials to the target user.

In some optional embodiments, the first obtaining module 3251 is further configured to obtain user information of the target user and material information of the candidate material; and inputting the user information and the material information into a plurality of shared networks to perform feature extraction to obtain a first user feature and a first material feature corresponding to each recommended target, wherein the plurality of shared networks are in one-to-one correspondence with the plurality of recommended targets.

In some optional embodiments, the prediction module 3252 is further configured to input the first user characteristic and the first material characteristic corresponding to each recommended target into a plurality of independent tower networks for predicting the recommended target, to obtain a predicted value of each recommended target, where the plurality of independent tower networks are in one-to-one correspondence with the plurality of recommended targets.

In some optional embodiments, the first obtaining module 3251 is further configured to obtain user information of the target user and material information of the candidate material; inputting the user information and the material information into a plurality of shared networks to perform feature extraction to obtain a second user feature and a second material feature corresponding to each recommended target, wherein the plurality of shared networks are in one-to-one correspondence with the plurality of recommended targets; and inputting the second user characteristics and the second material characteristics corresponding to each recommended target into a first sub-network in a plurality of independent tower networks for characteristic extraction to obtain the first user characteristics and the first material characteristics corresponding to each recommended target, wherein the plurality of independent tower networks are in one-to-one correspondence with the plurality of recommended targets.

In some alternative embodiments, each independent tower network includes a first subnetwork and a second subnetwork; the prediction module 3252 is further configured to input the first user characteristic and the first material characteristic output by the first sub-network of each independent tower network to a corresponding second sub-network to predict a recommended target, so as to obtain a predicted value of each recommended target.

In some optional embodiments, the apparatus further includes a first updating module, configured to obtain behavior data of the target user on the material to be recommended; according to the behavior data, respectively determining a first deviation value of a predicted value of each recommended target; based on the first deviation value of each recommended target, parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target are respectively updated.

In some optional embodiments, the updating module is further configured to weight and sum each first deviation value based on a weight of the first deviation value of the predicted value of each recommended target, to obtain a second deviation value; calculating a first gradient value based on the second deviation value; parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target are updated based on the first gradient values.

In some optional embodiments, the apparatus further includes a second updating module configured to determine a third deviation value of the fused predicted value through a preset loss function; calculating a second gradient value based on the third deviation value; based on the second gradient value, parameters of the fusion network are updated.

Fig. 7 is a schematic structural diagram of a server according to an embodiment of the present application, where the server 710 is configured to implement functions of a recommendation system. The server 71 may include one or more central processing units (central processing units, CPU) 711 (e.g., one or more processors) and memory 712, one or more storage media 713 (e.g., one or more mass storage devices) storing applications 721 or data 722, which vary in configuration or performance. Wherein the memory 712 and storage medium 713 may be transitory or persistent storage. The program stored on the storage medium 713 may include one or more modules (not shown), each of which may include a series of instruction operations on a computer device. Still further, the central processor 711 may be provided in communication with a storage medium 713, executing a series of instruction operations in the storage medium 713 on the computer device 710.

The server 710 may also include one or more power supplies 714, one or more wired or wireless network interfaces 715, one or more input/output interfaces 716, and/or one or more operating systems 717, such as Windows Server, mac OS XTM, uniTM, linuxT M, freeBSDTM, and the like.

Embodiments of the present application provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the recommendation method of multiple recommendation targets according to the embodiment of the application.

The present embodiments provide a computer-readable storage medium storing executable instructions or a computer program stored therein, which when executed by a processor, cause the processor to perform a recommendation method for multiple recommendation targets provided by the embodiments of the present application, for example, a recommendation method for multiple recommendation targets as shown in fig. 4 or 5.

In some embodiments, the computer readable storage medium may be FRAM, ROM, PROM, EP ROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; but may be a variety of devices including one or any combination of the above memories.

In some embodiments, the executable instructions may be in the form of programs, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and they may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

As an example, the executable instructions may, but need not, correspond to files in a file system, may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a hypertext markup language (H yper Text Markup Language, HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code).

As an example, executable instructions may be deployed to be executed on one computing device or on multiple computing devices located at one site or, alternatively, distributed across multiple sites and interconnected by a communication network.

In summary, based on the first user characteristics of the target user and the first material characteristics of the material to be recommended, the predicted values of the plurality of recommended targets are fused through the fusion network. In this way, in the process of fusing the predicted values of the plurality of recommended targets through the fusion network, the first user characteristics of the target users are considered, and the first material characteristics of the candidate materials are considered. Aiming at each target user, based on the difference of the first user characteristics and the first material characteristics, the fusion of the predicted values of the personalized multi-recommendation targets is realized, so that the probability that the materials recommended based on the fused predicted values meet the interest points of the target user is higher, and the recommendation accuracy is improved.

The foregoing is merely exemplary embodiments of the present application and is not intended to limit the scope of the present application. Any modifications, equivalent substitutions, improvements, etc. that are within the spirit and scope of the present application are intended to be included within the scope of the present application.

Claims

1. A material recommendation method based on multiple recommendation targets, the method comprising:

acquiring first user characteristics of a target user and first material characteristics of candidate materials;

according to the first user characteristics and the first material characteristics, respectively determining a predicted value of each recommended target in a plurality of recommended targets of the candidate material by the target user, wherein the predicted value of each recommended target is used for representing predicted information of user behaviors of the candidate material by the target user, and the user behaviors correspond to each recommended target;

inputting the first user characteristic, the first material characteristic and the predicted value of each recommended target into a fusion network to perform nonlinear weighted fusion of a plurality of predicted values, so as to obtain fusion predicted values of the plurality of recommended targets; and the fusion predicted value is used for recommending the candidate materials to the target user.

2. The method of claim 1, wherein the obtaining the first user characteristic of the target user and the first material characteristic of the candidate material comprises:

acquiring user information of the target user and material information of the candidate materials;

And inputting the user information and the material information into a plurality of sharing networks to perform feature extraction, so as to obtain the first user feature and the first material feature corresponding to each recommendation target, wherein the plurality of sharing networks are in one-to-one correspondence with the plurality of recommendation targets.

3. The method of claim 2, wherein determining a predicted value of the target user for each of a plurality of recommended targets of a candidate material based on the first user characteristic and the first material characteristic, respectively, comprises:

and inputting the first user characteristics and the first material characteristics corresponding to each recommended target into a plurality of independent tower networks to predict the recommended targets, and obtaining a predicted value of each recommended target, wherein the plurality of independent tower networks are in one-to-one correspondence with the plurality of recommended targets.

4. The method of claim 1, wherein the obtaining the first user characteristic of the target user and the first material characteristic of the candidate material comprises:

inputting the user information and the material information into a plurality of shared networks for feature extraction to obtain a second user feature and a second material feature corresponding to each recommended target, wherein the plurality of shared networks are in one-to-one correspondence with the plurality of recommended targets;

And inputting the second user characteristics and the second material characteristics corresponding to each recommended target into a first sub-network in a plurality of independent tower networks for characteristic extraction to obtain the first user characteristics and the first material characteristics corresponding to each recommended target, wherein the plurality of independent tower networks are in one-to-one correspondence with the plurality of recommended targets.

5. The method of claim 4, wherein determining a predicted value of the target user for each of a plurality of recommended targets of a candidate material based on the first user characteristic and the first material characteristic, respectively, comprises:

and inputting the first user characteristics and the first material characteristics output by the first sub-network of each independent tower network into a corresponding second sub-network to predict recommended targets, and obtaining a predicted value of each recommended target.

6. The method of claim 3 or 5, wherein after determining the predicted value of the target user for each of a plurality of recommended targets of a candidate material based on the first user characteristic and the first material characteristic, respectively, the method further comprises:

Acquiring behavior data of the target user on the candidate materials;

according to the behavior data, respectively determining a first deviation value of the predicted value of each recommended target;

and updating parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target based on the first deviation value of the predicted value of each recommended target.

7. The method of claim 6, wherein updating parameters of the shared network corresponding to each recommended target and/or the independent tower network corresponding to each recommended target based on the first deviation value of the predicted value of each recommended target comprises:

carrying out weighted summation on the first deviation value based on the weight of the first deviation value to obtain a second deviation value;

calculating a first gradient value based on the second deviation value;

and updating parameters of a shared network corresponding to each recommended target and/or an independent tower network corresponding to each recommended target based on the first gradient value.

8. The method of claim 1, wherein the inputting the first user characteristic, the first material characteristic, and the predicted value of each recommended target into a fusion network performs nonlinear weighted fusion of a plurality of predicted values, and after obtaining the fused predicted values, the method further comprises:

Determining a third deviation value of the fusion predicted value through a preset loss function;

calculating a second gradient value based on the third deviation value;

and updating parameters of the fusion network based on the second gradient value.

9. The method of claim 8, wherein the loss function satisfies the following expression:

wherein loss () is the third deviation value, prediction is the fusion predicted value, y _i Is the attribute value, mu, of the corresponding behavior of the ith recommended target _yi Is the mean value, sigma, of the attribute values of the corresponding behavior of the ith recommended target _yi Is the variance of the attribute values of the corresponding behavior of the ith recommendation target.

10. A multi-recommendation-target recommendation device, the device comprising:

the first acquisition module is used for acquiring first user characteristics of a target user and first material characteristics of candidate materials;

the prediction module is used for respectively determining the predicted value of each recommended target in a plurality of recommended targets of the target user to the candidate material according to the first user characteristics and the first material characteristics, wherein the predicted value of each recommended target is used for representing the predicted information of the target user on the user behavior of the candidate material, and the user behavior corresponds to each recommended target;

The fusion module is used for inputting the first user characteristics, the first material characteristics and the predicted values of each recommended target into a fusion network to carry out nonlinear weighted fusion of a plurality of predicted values, so as to obtain fusion predicted values of the plurality of recommended targets; and the fusion predicted value is used for recommending the candidate materials to the target user.

11. A multi-recommendation-target recommendation device, characterized in that the recommendation device comprises:

a memory for storing executable instructions or computer programs;

a processor for implementing the method according to any one of claims 1 to 9 when executing the executable instructions or computer programs stored in the memory.

12. A computer readable storage medium storing executable instructions or a computer program, which when executed by a processor, implements the method of any one of claims 1 to 9.