CN112163165A

CN112163165A - Information recommendation method, device, equipment and computer readable storage medium

Info

Publication number: CN112163165A
Application number: CN202011131347.1A
Authority: CN
Inventors: 张晗; 马连洋; 衡阵
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2020-10-21
Filing date: 2020-10-21
Publication date: 2021-01-01
Anticipated expiration: 2040-10-21
Also published as: CN112163165B

Abstract

The embodiment of the application provides an information recommendation method, an information recommendation device, information recommendation equipment and a computer readable storage medium, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring at least one first-class feature and at least one second-class feature of each piece of information to be recommended in an information set to be recommended, and performing feature cross processing on the at least one first-class feature to obtain a first prediction score; performing feature fusion processing on the at least one second type feature to obtain a second prediction score; performing prediction result transformation processing on the first prediction score and the second prediction score to obtain a tendency score of the information to be recommended; and recommending at least one piece of information to be recommended in the information set to be recommended to a target object according to the tendency score of each piece of information to be recommended. By the embodiment of the application, the popularity of the information to be recommended on the specific crowd can be accurately depicted, and the information recommendation effect on the specific crowd is improved.

Description

Information recommendation method, device, equipment and computer readable storage medium

Technical Field

The embodiment of the application relates to the technical field of internet, and relates to but is not limited to an information recommendation method, an information recommendation device, information recommendation equipment and a computer-readable storage medium.

Background

In an information flow recommendation scene, user characteristics and content characteristics are indispensable to a recommendation system, most of the existing content characteristics are defined based on prior information of the content, for example, videos are classified into sports and movies according to the video content, and are not linked with user behaviors, and for a user side, the recommendation system constructs user interest point characteristics corresponding to the content characteristics by using historical consumption behaviors of the user except for using basic information of the user. And when recommending, the click behavior prediction is carried out by utilizing the characteristics of the user and the video content. However, the recommendation thought has a great problem in new user recommendation, and since the new user does not have a history consumption behavior record, the recommendation system cannot acquire the interest point characteristics of the user, and can only perform recommendation by using the user basic information, so that the recommendation system cannot accurately predict the content of interest of the new user. Meanwhile, the recommendation of the new user is very important, and the retention of the new user determines key indexes of the whole recommended product, such as user scale and the like, so that the cold start problem of the new user is an important problem often faced by the existing recommendation system.

Aiming at the cold start problem of the new user, the technical scheme in the related art can only recommend by using the basic attribute of the user. For example, local news content is recommended based on the user's regional information; or based on the information of the gender, the age and the like of the user, the consumption content of the crowd under the specific gender and the age is counted, and the high consumption content is recommended.

However, the solution in the related art is only applicable to the content which has already been consumed, the newly produced content cannot be counted, and the recommendation system may have a poor recommendation effect for a specific population due to the complexity of the recommendation system.

Disclosure of Invention

The embodiment of the application provides an information recommendation method, an information recommendation device, information recommendation equipment and a computer readable storage medium, and relates to the technical field of artificial intelligence. The method comprises the steps of respectively carrying out feature cross processing and feature fusion processing on at least one first class feature and at least one second class feature of information to be recommended, carrying out prediction result transformation processing on the obtained first prediction score and the obtained second prediction score to obtain the tendency score of each piece of information to be recommended, and carrying out information recommendation according to the tendency score.

The technical scheme of the embodiment of the application is realized as follows:

the embodiment of the application provides an information recommendation method, which comprises the following steps:

acquiring at least one first type feature and at least one second type feature of each piece of information to be recommended in an information set to be recommended; performing feature cross processing on the at least one first type feature to obtain a first prediction score; performing feature fusion processing on the at least one second type feature to obtain a second prediction score; performing prediction result transformation processing on the first prediction score and the second prediction score to obtain a tendency score of the information to be recommended; and recommending at least one piece of information to be recommended in the information set to be recommended to a target object according to the tendency score of each piece of information to be recommended.

An embodiment of the present application provides an information recommendation device, including:

the acquisition module is used for acquiring at least one first type feature and at least one second type feature of each piece of information to be recommended in the information set to be recommended; the characteristic cross processing module is used for carrying out characteristic cross processing on the at least one first type characteristic to obtain a first prediction score; the characteristic fusion processing module is used for carrying out characteristic fusion processing on the at least one second type characteristic to obtain a second prediction score; the transformation processing module is used for carrying out transformation processing on the prediction results of the first prediction score and the second prediction score to obtain the tendency score of the information to be recommended; and the recommending module is used for recommending at least one piece of information to be recommended in the information set to be recommended to a target object according to the tendency score of each piece of information to be recommended.

Embodiments of the present application provide a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium; the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor is configured to execute the computer instructions to implement the information recommendation method.

An embodiment of the present application provides an information recommendation device, including: a memory for storing executable instructions; and the processor is used for realizing the information recommendation method when executing the executable instructions stored in the memory.

The embodiment of the application provides a computer-readable storage medium, which stores executable instructions for causing a processor to execute the executable instructions to implement the information recommendation method.

The embodiment of the application has the following beneficial effects: aiming at each piece of information to be recommended in the information set to be recommended, feature cross processing and feature fusion processing are respectively carried out on at least one first type feature and at least one second type feature, a first prediction score and a second prediction score are correspondingly obtained, prediction results of the first prediction score and the second prediction score are converted, and a tendency score of each piece of information to be recommended is obtained. Therefore, the final tendency score is determined based on the first class characteristic and the second class characteristic of the information to be recommended respectively, so that the complex diversity of the characteristics of the information to be recommended is fully considered, the popularity of the information to be recommended on the specific crowd can be accurately described, and the information recommendation effect on the specific crowd is improved.

Drawings

FIG. 1 is an alternative architecture diagram of an information recommendation system provided by an embodiment of the present application;

FIG. 2 is a schematic structural diagram of a server provided in an embodiment of the present application;

fig. 3 is an alternative flow chart of an information recommendation method provided in an embodiment of the present application;

FIG. 4 is a schematic flow chart of an alternative information recommendation method provided in the embodiments of the present application;

FIG. 5 is an alternative flow chart of an information recommendation method provided in an embodiment of the present application;

FIG. 6 is an alternative flow chart of an information recommendation method provided in an embodiment of the present application;

FIG. 7 is an alternative flow chart of an information recommendation method provided in an embodiment of the present application;

FIG. 8 is an alternative flow chart of an information recommendation method provided in an embodiment of the present application;

fig. 9A is a block diagram of a tendency recognition network provided in an embodiment of the present application;

fig. 9B is an alternative flowchart of a training method for a tendency recognition network according to an embodiment of the present application;

fig. 10 is a network configuration diagram of the overall algorithm in the embodiment of the present application.

Detailed Description

In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be described in further detail with reference to the attached drawings, the described embodiments should not be considered as limiting the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict. 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 the embodiments of the present application belong. The terminology used in the embodiments of the present application is for the purpose of describing the embodiments of the present application only and is not intended to be limiting of the present application.

In order to better understand the network structure search method provided in the embodiment of the present application, first, a network structure search method in the related art is explained:

in the related art, only basic attributes of a user can be used for recommending a cold start problem of a new user. However, this solution is only applicable to the content that has already been consumed, and for the newly produced content, statistics cannot be performed. Therefore, the application provides a crowd tendency label of the content aiming at the pain point, and predicts the popularity of the content to a specific crowd by only utilizing the characteristics of the content. Before the content is recommended, the user identification of the content which is popular can be provided for the recommendation system, the recommendation system is guided to recommend the new user under the specific crowd, and the cold start problem of the new user is relieved.

In addition, the complexity of the recommendation system may lead to poor recommendation for a certain group of users, for example, in the case of a small number of female users, the recommendation system may not be able to learn the click behavior of the group of users well. Therefore, in a scene of optimizing the recommendation effect for a specific crowd, the estimation of the crowd tendency is also necessary, for example, the method can assist in operation to perform content screening of the specific crowd tendency and the like.

Based on the needs and problem analysis of the related art, the present application addresses the following issues: firstly, defining the group popularity of content (namely information to be recommended) based on consumption indexes to be optimized of a recommendation system, such as click rate, Page View (PV), duration, sharing and the like. If the optimization is carried out based on the click rate, dividing the video content consumed by the group into a popular type and an unpopular type according to the click rate; secondly, constructing a wide & deep classification model by using the constructed classification data and using multi-modal characteristics of the title, classification, account number, label, cover, video frame and the like of the video; and finally, predicting the content by using the trained model, and using the tendency score output by the last layer of the classification network as the tendency score of the content on the crowd.

The information recommendation method provided by the embodiment of the application comprises the steps of firstly, obtaining at least one first-class feature and at least one second-class feature of each piece of information to be recommended in an information set to be recommended; then, performing feature cross processing on at least one first type feature to obtain a first prediction score; performing feature fusion processing on at least one second type feature to obtain a second prediction score; performing prediction result transformation processing on the first prediction score and the second prediction score to obtain a tendency score of the information to be recommended; and finally, recommending at least one piece of information to be recommended in the information set to be recommended to the target object according to the tendency score of each piece of information to be recommended. Therefore, the final tendency score is determined based on the first class characteristic and the second class characteristic of the information to be recommended respectively, so that the complex diversity of the characteristics of the information to be recommended is fully considered, the popularity of the information to be recommended on the specific crowd can be accurately depicted, and the information recommendation effect on the specific crowd is improved.

An exemplary application of the information recommendation device according to the embodiment of the present application is described below, in one implementation, the information recommendation device according to the embodiment of the present application may be implemented as any terminal such as a notebook computer, a tablet computer, a desktop computer, a mobile device (e.g., a mobile phone, a portable music player, a personal digital assistant, a dedicated messaging device, a portable game device), an intelligent robot, and in another implementation, the information recommendation device according to the embodiment of the present application may also be implemented as a server. In the following, an exemplary application when the information recommendation is implemented as a server will be explained.

Referring to fig. 1, fig. 1 is an alternative architecture diagram of an information recommendation system 10 provided in an embodiment of the present application. In the embodiment of the present application, information to be recommended is taken as a video for example, in order to recommend a generated video to a user who is more interested in the video or a new user, an information recommendation system 10 provided in the embodiment of the present application includes a terminal 100, a network 200, and a server 300, where a video recommendation application is run on the terminal 100, a user may register on the video recommendation application to become a new user, after successful registration, the video recommendation application may actively recommend a video to the user, and the recommended video is a video that the user may be interested in, then the terminal 100 may send a registration completion message to the server 300 through the network 200 to notify that the current user is a new user, and the server 300 obtains at least one first-class feature and at least one second-class feature of each video to be recommended in a set of videos to be recommended; then, performing feature cross processing on at least one first type feature to obtain a first prediction score; performing feature fusion processing on at least one second type feature to obtain a second prediction score; performing prediction result transformation processing on the first prediction score and the second prediction score to obtain a tendency score of the video to be recommended; according to the tendency score of each video to be recommended, at least one video to be recommended is determined in the video set to be recommended, and the videos to be recommended are recommended to the terminal 100.

The information recommendation method provided by the embodiment of the application further relates to the technical field of artificial intelligence, and can be at least realized through machine learning and natural language processing technologies in the artificial intelligence technology. Machine Learning (ML) is a one-field multi-field cross subject, and relates to multiple subjects such as probability theory, statistics, approximation theory, convex analysis and algorithm complexity theory. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formal education learning. In the embodiment of the application, the response to the network structure search request is realized through a machine learning technology so as to automatically search a target network structure, and the training and model optimization of the controller and the score model are realized. Natural Language Processing (NLP) is an important direction in the fields of computer science and artificial intelligence, and it is studying various theories and methods that can achieve effective communication between people and computers using natural Language. Natural language processing is a science integrating linguistics, computer science and mathematics. Therefore, the research in this field will involve natural language, i.e. the language that people use everyday, so it is closely related to the research of linguistics. Natural language processing techniques typically include text processing, semantic understanding, machine translation, robotic question and answer, knowledge mapping, and the like.

Fig. 2 is a schematic structural diagram of a server 300 according to an embodiment of the present application, where the server 300 shown in fig. 2 includes: at least one processor 310, memory 350, at least one network interface 320, and a user interface 330. The various components in server 300 are coupled together by a bus system 340. It will be appreciated that the bus system 340 is used to enable communications among the components connected. The bus system 340 includes a power bus, a control bus, and a status signal bus in addition to a data bus. For clarity of illustration, however, the various buses are labeled as bus system 340 in fig. 2.

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

The user interface 330 includes one or more output devices 331, including one or more speakers and/or one or more visual display screens, that enable presentation of media content. The user interface 330 also includes one or more input devices 332, including user interface components to facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

The memory 350 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid state memory, hard disk drives, optical disk drives, and the like. Memory 350 optionally includes one or more storage devices physically located remote from processor 310. The memory 350 may include either volatile memory or nonvolatile memory, and may also include both volatile and nonvolatile memory. The nonvolatile Memory may be a Read Only Memory (ROM), and the volatile Memory may be a Random Access Memory (RAM). The memory 350 described in embodiments herein is intended to comprise any suitable type of memory. In some embodiments, memory 350 is capable of storing data, examples of which include programs, modules, and data structures, or subsets or supersets thereof, as exemplified below, to support various operations.

An operating system 351 including system programs for processing 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 processing hardware-based tasks;

a network communication module 352 for communicating to other computing devices via one or more (wired or wireless) network interfaces 320, exemplary network interfaces 320 including: bluetooth, wireless compatibility authentication (WiFi), and Universal Serial Bus (USB), etc.;

an input processing module 353 for detecting one or more user inputs or interactions from one of the one or more input devices 332 and translating the detected inputs or interactions.

In some embodiments, the apparatus provided by the embodiments of the present application may be implemented in software, and fig. 2 illustrates an information recommendation apparatus 354 stored in the memory 350, where the information recommendation apparatus 354 may be an information recommendation apparatus in the server 300, and may be software in the form of programs and plug-ins, and the like, and includes the following software modules: the acquisition module 3541, the feature intersection processing module 3542, the feature fusion processing module 3543, the transformation processing module 3544, and the recommendation module 3545 are logical and thus may be arbitrarily combined or further separated depending on the functions implemented. The functions of the respective modules will be explained below.

In other embodiments, the apparatus provided in the embodiments of the present Application may be implemented in hardware, and for example, the apparatus provided in the embodiments of the present Application may be a processor in the form of a hardware decoding processor, which is programmed to execute the information recommendation method provided in the embodiments of the present Application, for example, the processor in the form of the hardware decoding processor may be one or more Application Specific Integrated Circuits (ASICs), DSPs, Programmable Logic Devices (PLDs), Complex Programmable Logic Devices (CPLDs), Field Programmable Gate Arrays (FPGAs), or other electronic components.

The information recommendation method provided by the embodiment of the present application will be described below in conjunction with an exemplary application and implementation of the server 300 provided by the embodiment of the present application. Referring to fig. 3, fig. 3 is an alternative flowchart of an information recommendation method provided in an embodiment of the present application, and will be described with reference to the steps shown in fig. 3.

Step S301, at least one first-class feature and at least one second-class feature of each piece of information to be recommended in the information set to be recommended are obtained.

Here, the information set to be recommended includes at least one piece of information to be recommended, the information to be recommended in the information set to be recommended may have the same category, for example, all of the information to be recommended is video information, or all of the information to be recommended is text information, and the information to be recommended in the information set to be recommended may have a certain correlation, for example, an interactive relationship is provided between each of the information to be recommended in the information set to be recommended and a specific target object, that is, there is interactive data between each of the information to be recommended in the information set to be recommended and any user group, and a user in the user group clicks or browses each of the information to be recommended in the information set to be recommended, or each of the information to be recommended in the information set to be recommended is recommended to a.

In the embodiment of the application, when information is recommended, at least one piece of recommendation information in the information set to be recommended is recommended to a target user, so that each piece of information to be recommended in the information set to be recommended needs to be analyzed to determine whether the information to be recommended can be recommended or not.

In the embodiment of the application, when each piece of information to be recommended is analyzed, the analysis is performed based on the first type of features and the second type of features of the information to be recommended. For example, the first type of features may be low-order features, and the second type of features may be high-order features, where the low-order features are features calculated or determined from fewer original features in the information to be recommended, and the high-order features are features calculated or determined from more original features in the information to be recommended.

For example, when the information to be recommended is a video, the first type of feature may be a low-order feature that can be determined by a first number of original features of the video, such as a category of the video, a video duration, or a video quality score, and the second type of feature may be a high-order feature that can be determined by a second number of original features of the video, such as a title, a cover, or a label of the video.

In some embodiments, the number of the original features of the information to be recommended, which are related to the first type of features, may also be smaller than the number of the original features of the information to be recommended, for example, the first type of features need to be determined by a first number of the original features in the information to be recommended, and the second type of features need to be determined by a second number of the original features in the information to be recommended, so at this time, the first number may be smaller than the second number. Of course, the number of the original features of the information to be recommended related to the first type of features may also be greater than or equal to the number of the original features of the information to be recommended related to the second type of features, which may be specifically determined according to input information of a network structure in the information recommendation model, and the embodiment of the present application is not particularly limited.

Step S302, performing feature cross processing on at least one first-class feature to obtain a first prediction score.

Here, after the first class features are obtained, feature crossing processing is performed on the first class features, where the feature crossing processing may be implemented by: firstly, each first-class feature is subjected to feature embedding processing to obtain an embedded vector, then, linear transformation is carried out on the embedded vectors corresponding to different first-class features, for example, calculation such as vector multiplication or vector summation can be carried out, so that feature cross processing is carried out on the embedded vectors corresponding to a plurality of first-class features to obtain a final output result, wherein the output result is a one-dimensional vector, and the output one-dimensional vector is a first prediction score.

Step S303, performing feature fusion processing on at least one second type feature to obtain a second prediction score.

Here, after the second type of feature is obtained, feature fusion processing is performed on the second type of feature, where the feature fusion processing may be implemented by: since the second type of features may include at least a text type feature and an image type feature, the text type feature and the image type feature may be processed separately. Firstly, performing word embedding processing on text features in the second type of features to obtain a word embedding vector of each word, and then solving the average value of all word embedding vectors, so that the encoding process of the text features in the second type of features is realized; for the image features in the second features, a specific feature extraction network can be directly adopted for feature extraction to obtain vectors after the image features are coded; then, the vectors after the encoding of the text-type features and the image-type features are respectively subjected to weighting processing and linear transformation, then splicing and fusion are carried out, finally, the dimension of the spliced vectors is reduced to one dimension through the linear transformation, and a one-dimensional vector is obtained, wherein the one-dimensional vector is the second prediction score.

In some embodiments, when the text-class feature included in the second-class feature is composed of a plurality of words, for example, when the text-class feature is a sentence or an article, since the sentence or the article is a text composed of a plurality of words, before performing the feature fusion process, the sentence or the article needs to be subjected to the word segmentation process first, and then the word embedding process is performed.

In the embodiment of the present application, the text-type feature may include a feature composed of a plurality of words, such as a title and a body, and may further include a tag feature composed of one word. The word segmentation processing can be carried out only on the characteristics of a plurality of words, such as a title, a text and the like, and then the word embedding processing is carried out on each word obtained after the word segmentation processing to obtain a word embedding vector; because the label is usually a single word, word segmentation processing is not carried out on the label characteristic, and word embedding processing can be directly carried out to obtain a word embedding vector.

And step S304, performing prediction result transformation processing on the first prediction score and the second prediction score to obtain the tendency score of the information to be recommended.

And integrating the first prediction score and the second prediction score, and summing and transforming the first prediction score and the second prediction score to obtain the tendency score of the information to be recommended.

Step S305, recommending at least one piece of information to be recommended in the information set to be recommended to the target object according to the tendency score of each piece of information to be recommended.

In the embodiment of the application, information to be recommended, of which the tendency score is greater than the threshold value, may be recommended to the target object, for example, the value of the tendency score may be any value between 0 and 1, and the threshold value may be 0.5; or, the information to be recommended in the information set to be recommended may be sorted according to the descending order of the tendency scores to form an information sequence to be recommended, and then a specific number of information to be recommended before the information sequence to be recommended are recommended to the target object in sequence according to the order of the information sequence to be recommended.

The target object may be an object belonging to a particular group, for example, the target object may be a new user, or may be a female group user.

The information recommendation method provided by the embodiment of the application is used for respectively performing feature cross processing and feature fusion processing on at least one first-class feature and at least one second-class feature aiming at each piece of information to be recommended in an information set to be recommended, correspondingly obtaining a first prediction score and a second prediction score, and performing prediction result conversion processing on the first prediction score and the second prediction score to obtain the tendency score of each piece of information to be recommended, so that at least one piece of information to be recommended in the information set to be recommended is recommended according to the tendency score of each piece of information to be recommended, namely, the information to be recommended is recommended to a target object belonging to a specific group. Therefore, the final tendency score is determined based on the first class characteristic and the second class characteristic of the information to be recommended respectively, so that the complex diversity of the characteristics of the information to be recommended is fully considered, the popularity of the information to be recommended on the specific crowd can be accurately described, and the information recommendation effect on the specific crowd is improved.

In some embodiments, the method of the embodiment of the present application may also be applied to an initial recommendation process after a new user completes registration, where the information recommendation system includes a terminal and a server, the user may register on the terminal to become a new user of any application, and after the user registration is completed, the application performs information recommendation to the user to further obtain a consumption log of the user, where the consumption log includes but is not limited to: click frequency, viewing duration, exposure times and the like.

Fig. 4 is an optional flowchart schematic diagram of an information recommendation method provided in an embodiment of the present application, and as shown in fig. 4, the method includes the following steps:

in step S401, the terminal receives a registration request from a user.

Step S402, after the user finishes registering at the client of the specific user on the terminal, the terminal sends an information recommendation request to the server, wherein the information recommendation request comprises the basic information of the user.

Here, the basic information of the user is information input in the user registration process, for example, information of contact address, sex, age, current location, and the like.

In step S403, the server obtains at least one first-class feature and at least one second-class feature of each piece of information to be recommended in the information set to be recommended.

Step S404, the server carries out feature cross processing on at least one first type feature to obtain a first prediction score.

Step S405, the server performs feature fusion processing on at least one second type feature to obtain a second prediction score.

In step S406, the server performs prediction result transformation processing on the first prediction score and the second prediction score to obtain a tendency score of the information to be recommended.

Step S407, the server selects at least one piece of information to be recommended from the information set to be recommended according to the tendency score of each piece of information to be recommended.

It should be noted that steps S403 to S403 are the same as steps S301 to S305, and the embodiment of the present application is not repeated.

And step S408, the server recommends the selected at least one piece of information to be recommended to the terminal.

According to the information recommendation method provided by the embodiment of the application, after the new user finishes registration, because the new user does not have interactive information in the application, the interest point information of the user cannot be determined according to the use habit of the new user, so that the information to be recommended with a high tendency score is selected from the information set to be recommended by adopting the method provided by the embodiment of the application, the information is information which the new user may be interested in, and the information is recommended to the new user, so that the accuracy of initial recommendation can be improved, and the recommendation effect is improved.

In some embodiments, before information recommendation is performed, an information set to be recommended needs to be determined, so that when information recommendation is performed, at least one piece of information to be recommended may be determined from the information set to be recommended, based on fig. 3, fig. 5 is an optional flowchart of the information recommendation method provided in the embodiment of the present application, and as shown in fig. 5, before step S301, the method further includes the following steps:

in step S501, attribute information of the target object is determined. The attribute information of the target object may be information such as a type or a characteristic of the target object, for example, the attribute information of the target object may be that the target object is a female user, that the target object is between 20 and 30 years old, or the like.

Step S502, an object set formed by the objects with the attribute information and a recommendation information set formed by recommendation information with interactive information between the objects in the object set are obtained.

Here, each object in the set of objects has the attribute information, for example, the objects in the set of objects may all be female users; each piece of recommendation information in the recommendation information set and any object in the object set have interaction information therebetween, and the interaction information may be information corresponding to any one of interaction operations such as clicking, browsing, purchasing, and sharing, for example, information purchased by female users in the female user set may be used as recommendation information to form the recommendation information set.

Step S503, for each piece of recommendation information in the recommendation information set, acquiring interaction data between each object in the object set and the recommendation information.

Here, taking the interaction data as click data as an example, the number of clicks of each piece of recommendation information by each female user in the female user set is obtained.

Step S504, the interaction data are aggregated, and the interaction probability of each recommendation information for the object in the object set is obtained.

Here, the click frequency of all female users is summed up, and then averaged in the female user set, so as to obtain the click probability of the female user set for each piece of recommendation information.

And step S505, selecting at least one piece of recommendation information from the recommendation information set as information to be recommended according to the interaction probability to form an information set to be recommended.

Here, the recommendation information with the interaction probability greater than the probability threshold is selected as the information to be recommended, or a preset number of recommendation information is selected as the information to be recommended according to the sequence of the interaction probability from large to small.

In some embodiments, the interaction data includes at least exposure times, click times and browsing duration of the recommendation information; based on fig. 5, fig. 6 is an optional flowchart of the information recommendation method provided in the embodiment of the present application, and as shown in fig. 6, step S504 may be implemented by the following steps:

step S601, aggregating the exposure times, the click times and the browsing duration to obtain an average click rate and an average browsing duration for each piece of recommended information.

Here, the average click rate is an average value of the total number of clicks corresponding to the objects in the object set over the entire object set, that is, a value obtained by dividing the total number of clicks by the number of objects in the object set; the average browsing duration is an average value of the total browsing duration corresponding to the objects in the object set over the entire object set, that is, a value obtained by dividing the total browsing duration by the number of the objects in the object set.

Step S602, determining any one of the average click rate and the average browsing duration as the interaction probability of the recommendation information.

Referring to fig. 6, in some embodiments, the method further includes the following steps:

step S603, determining a total average click rate corresponding to all objects in the object set according to the average click rate of each piece of recommendation information in the recommendation information set.

Here, the total average click rate refers to an average value of the average click rate of each piece of recommendation information in the entire recommendation information set, that is, a value obtained by dividing the sum of the average click rates corresponding to all objects by the number of pieces of recommendation information in the recommendation information set.

Step S604, determining the total average browsing duration corresponding to all the objects in the object set according to the average browsing duration of each piece of recommended information in the recommended information set.

Here, the total average browsing duration refers to an average value of the average browsing duration of each piece of recommendation information in the entire recommendation information set, that is, a value obtained by dividing the sum of the average browsing durations corresponding to all the objects by the number of pieces of recommendation information in the recommendation information set.

Referring to fig. 6, in some embodiments, step S505 can be implemented by any one of the following ways: the first method is as follows: step S605, selecting the recommendation information with the average click rate larger than the total average click rate as the information to be recommended from the recommendation information set.

Square two: step S606, selecting the recommendation information with the average browsing time length larger than the total average browsing time length from the recommendation information set as the information to be recommended.

In some embodiments, the first type of feature comprises at least one of an information category feature, an information length feature, and an information quality feature; when the first type of features only comprise the information type features, performing feature cross processing based on the information type features to obtain a first prediction score; when the first type of features only comprise information length features, performing feature cross processing based on the information length features to obtain a first prediction score; and when the first type of features only comprise the information quality features, performing feature cross processing based on the information quality features to obtain a first prediction score. Of course, the first type of feature may also include an information category feature, an information length feature, and an information quality feature, and then feature cross processing may be performed based on the information category feature, the information length feature, and the information quality feature to obtain the first prediction score. The first type of features including an information type feature, an information length feature, and an information quality feature are described as an example.

Based on fig. 3, fig. 7 is an optional flowchart of the information recommendation method provided in the embodiment of the present application, and as shown in fig. 7, step S302 may be implemented by the following steps:

step S701, respectively encoding the first-level category characteristics and the second-level category characteristics in the information category characteristics to correspondingly obtain first-level category characteristic vectors and second-level category characteristic vectors.

Here, the information category feature refers to that the information to be recommended is assigned to a certain category according to the content of the information to be recommended, the categories may be a primary category and a secondary category, the primary category is a parent of the secondary category, and each piece of information to be recommended may have not only one primary category but also one secondary category.

When the information category characteristics of the information to be recommended are acquired, the first-level category characteristics and the second-level category characteristics of the information to be recommended are acquired, so that the first-level category characteristics and the second-level category characteristics are encoded respectively to obtain a first-level category characteristic vector and a second-level category characteristic vector. In some embodiments, the encoding process may use a one-hot encoding scheme for encoding.

Step S702, carrying out discretization processing and coding processing on the information length characteristics in sequence to obtain information length characteristic vectors.

Since the information length feature may be a continuous feature, the continuous information length feature may be discretized, for example, by using equal-frequency binning discretization to obtain a plurality of discretized levels, and then, the discretized plurality of features may be encoded to obtain the information length feature vector.

Step S703, encoding the information quality characteristic to obtain an information quality characteristic vector.

In some embodiments, step S703 may also be implemented by:

step S7031, a pre-scoring value for the information to be recommended is acquired.

Step S7032, coding the pre-scored value to obtain an information quality characteristic vector.

In the embodiment of the application, the information quality can be scored firstly to obtain the pre-scoring value of the information to be recommended, namely, the information to be recommended can be scored according to at least one prior standard such as the authenticity of the information, the sufficiency of the content and the fineness of the content, then the scoring result is divided into a plurality of grades, and the plurality of grades are coded to obtain the information quality characteristic vector.

Step S704, a factor decomposition machine is adopted to carry out cross processing on the primary class characteristic vector, the secondary class characteristic vector, the information length characteristic vector and the information quality characteristic vector to obtain a first prediction score.

Here, the factorizer is configured to perform linear prediction on the first-level class feature vector, the second-level class feature vector, the information length feature vector, and the information quality feature vector, and finally output a one-dimensional vector, which is a first prediction score.

It should be noted that, the step of determining the first prediction score may also be implemented by using a wide network in the wide & deep network.

In some embodiments, the second type of feature includes at least a text feature and an image feature; referring to fig. 7, step S303 can be implemented by the following steps:

step S705, determining at least one word according to the text feature.

Here, when the text feature is a title or a body text, at least one word may be obtained by performing word segmentation processing on the title or the body text; when the text feature is a label, the word corresponding to the label can be directly acquired.

Step S706, a word embedding vector for each word is obtained.

In step S707, the average value of the word embedding vectors is determined as the encoding vector of the text feature.

Step S708, performing image feature extraction on the image features to obtain image feature vectors.

Step S709, based on the attention mechanism, performs feature fusion processing on the coding vector and the image feature vector to obtain a second prediction score.

Referring to fig. 7, in some embodiments, step S304 can be implemented by:

step S710, summing the first prediction score and the second prediction score to obtain a total prediction score.

And step S711, performing prediction result transformation processing on the total prediction score by adopting a nonlinear activation function to obtain the tendency score of the information to be recommended.

Based on fig. 7, fig. 8 is an optional flowchart of the information recommendation method provided in the embodiment of the present application, and as shown in fig. 8, step S709 may be implemented by:

in step S801, linear transformation processing is performed on the encoded vector and the image feature vector based on the attention mechanism, and the encoded vector after linear transformation and the image feature vector after linear transformation are obtained correspondingly.

In some embodiments, step S801 may also be implemented by:

in step S8011, a first attention weight of the encoded vector and a second attention weight of the image feature vector are determined. In step S8012, the coding vector and the image feature vector are weighted by the first attention weight and the second attention weight, respectively. The weighting here means that the first attention weight is multiplied by the coding vector and the second attention weight is multiplied by the image feature vector.

Step S8013, linear transformation processing is performed on the weighted encoding vector and the weighted image feature vector, and a linearly transformed encoding vector and a linearly transformed image feature vector are obtained correspondingly.

Here, FC linear transformation is performed on the weighted encoded vector and the weighted image feature vector using a full-Connected network (FC).

And S802, splicing the coding vector after the linear transformation and the image characteristic vector after the linear transformation to obtain a spliced vector.

Step S803, performing activation processing on the splicing vector to reduce the dimension of the splicing vector to one dimension, and obtaining a second prediction score. Here, the relu function may be used to activate the splicing vector, and after the relu function is activated, a one-dimensional vector is connected to reduce the dimension of the activated splicing vector to one dimension, so as to obtain the second prediction score.

It should be noted that, the step of determining the second prediction score may also be implemented by using a deep network in the wide & deep network.

In some embodiments, a tendency recognition network may be trained in advance, and the step of determining the tendency score of the information to be recommended in any of the above embodiments is implemented by the trained tendency recognition network. Based on this, an orientation recognition network and a training method of the orientation recognition network are provided in an embodiment of the present application, fig. 9A is a structural diagram of the orientation recognition network provided in the embodiment of the present application, and as shown in fig. 9A, an orientation recognition network 90 includes a feature intersection network 901, a feature fusion network 902, and a transformation processing layer 903, where the feature intersection network 901 is configured to perform feature intersection processing on a first type of features of input information to be recommended to obtain a first prediction score; the feature fusion network 902 is configured to perform feature fusion processing on a second type of features of the input information to be recommended to obtain a second prediction score; the transformation processing layer 903 is configured to perform transformation processing on the prediction result on the first prediction score output by the feature crossing network 901 and the second prediction score output by the feature fusion network 902 to obtain a tendency score of the information to be recommended.

Fig. 9B is an alternative flowchart of a training method for a tendency recognition network according to an embodiment of the present application, and as shown in fig. 9B, the training method includes the following steps:

step S901, inputting the first type sample features of the sample information and the second type sample features of the sample information into the tendency identification network.

And step S902, performing characteristic cross processing on the first type of sample characteristics through a characteristic cross network of the tendency identification network to obtain a first sample prediction score.

And step S903, performing characteristic fusion processing on the second type sample characteristics through a characteristic fusion network of the tendency identification network to obtain a second sample prediction score.

And step S904, performing prediction result transformation processing on the first sample prediction score and the second sample prediction score through a transformation processing layer of the tendency identification network to obtain a sample tendency score of the sample information.

Step S905, the sample tendency score is input into a preset loss model, and a loss result is obtained.

Here, the preset loss model is configured to compare the sample tendency score with a preset tendency score to obtain a loss result, where the preset tendency score may be a real tendency score preset by a user and corresponding to the sample information, and the preset tendency score may take a value of 0 or 1.

In this embodiment, the preset loss model includes a loss function, a distance between the sample tendency score and the preset tendency score is calculated through the loss function, and the loss result is determined according to the distance. When the distance between the sample tendency score and the preset tendency score is larger, the difference between the output result of the model and the true value is larger, and further training is needed; when the distance between the sample tendency score and the preset tendency score is smaller, the difference between the training result of the model and the true value is smaller, the difference between the output result of the model and the true value is smaller, the result is closer to the true value, and further training of the model can be considered to be stopped.

And step S906, according to the loss result, correcting parameters in the feature cross network and the feature fusion network to obtain a trained tendency recognition network.

Here, when the distance is smaller than the distance threshold, the loss result indicates that the current feature crossing network cannot perform accurate feature crossing processing on the first type of sample features, or indicates that the current feature fusion network cannot perform accurate feature fusion processing on the second type of sample features, and therefore, it is necessary to modify parameters in at least one of the feature crossing network and the feature fusion network until the distance between the sample tendency score output by the tendency recognition network and the preset tendency score is smaller than the distance threshold, and training of the tendency recognition network is stopped.

According to the training method for the tendency recognition network, the sample information is input into the tendency recognition network, the characteristic cross processing, the characteristic fusion processing and the prediction result transformation processing are sequentially performed on the sample information through the characteristic cross network and the characteristic fusion network, so that the sample tendency score of the sample information is obtained, and the sample tendency score is input into the preset loss model, so that the loss result is obtained. Therefore, parameters in the feature cross network and the feature fusion network can be corrected according to the loss result, and the obtained trained tendency recognition network can accurately predict the tendency score of the information to be recommended, so that accurate information recommendation is performed on the user.

In some embodiments, when the tendency recognition network is used for model prediction, each piece of information to be recommended may be sequentially input into the tendency recognition network, and the tendency recognition network may automatically output a tendency score corresponding to the piece of information to be recommended, where the tendency score may be any value between 0 and 1, and when the tendency score is greater than 0.5, it indicates that the popularity of the corresponding piece of information to be recommended is high, the piece of information to be recommended may be recommended to a target user; and when the tendency score is less than or equal to 0.5, indicating that the popularity of the corresponding information to be recommended is low, not recommending the information to be recommended to the target user.

Next, an exemplary application of the embodiment of the present application in a practical application scenario will be described.

The embodiment of the application provides an information recommendation method, which is a video crowd tendency estimation scheme based on multi-mode information and aims to score the popularity of video content on a specific crowd so as to improve the accuracy of personalized recommendation on the specific crowd. Firstly, modeling a concept of crowd tendency, dividing contents into likes and dislikes according to consumption indexes such as click rate, duration and the like by using consumption logs of crowds, namely converting the popularity of the contents into a binary classification problem; secondly, constructing a wide & deep classification model to fit the problems by utilizing the multi-modal information of the video, such as titles, classifications, labels, covers, video frames and the like; finally, in prediction, the output probability of the classification model is used as a popularity score (i.e., a tendency score) of the content over a specified population. On the recommending side, the score can be used for optimizing the recommending strategy of the specified crowd, for example, in the cold start stage of a new user, the recommending weight of the high-popularity score video is increased, and further the consumption indexes of clicking, duration, retention and the like of the specific crowd are improved.

The tendency data set (i.e. the information set to be recommended) used in the embodiment of the application is from a user consumption log in a main feed scene of an information flow product, wherein the data format of the user consumption log is (t, uid)_i，vid_v，exp_i,v，clk_i,v，playdur_i,v) Wherein t is a reporting timestamp, uid_iFor users i, vid_vFor video v, exp_i,vNumber of exposures of video v to user i, clk_i,vFor the number of clicks of user i on video v, playdur_i,vThe viewing duration of video v for user i. And screening the logs according to a designated user group (assumed as a female group), and only keeping records corresponding to female users. Define a certain consumption interval according to vid_vGrouping, aggregating the consumption data of exposure, click, duration and the like to obtain an aggregated data structure of (vid)_v，exp_v，clk_v，ctr_v，playdur_v) Wherein the total exposure number of the female of the video v is

Total number of clicks of video v is

Average click rate of video v is ctr_v＝clk_vexp_vThe average playing time of the video v is

U is the set of all female users. Further, the average click rate of the female population for the total content consumption can be found to be

The average duration of consumption of the entire content by the female population is

The female group corresponds to the object set, and all the contents correspond to the recommendation information set.

In the embodiment of the application, a video set which is popular with women or not can be constructed according to a service target (assumed as ctr) to be optimized. If it is

The video v is a positive sample, i.e. popular with women; otherwise, the sample is negative, i.e. not popular with women. The data set (i.e. the set of information to be recommended) required for classifying female trends is obtained.

Because there are many factors that influence whether video content is popular with a group, in addition to the title and cover features that are directly presented to the user, features such as video theme, video quality, etc. also have a direct effect on the result. Therefore, the embodiment of the application provides a wide & deep classification network structure based on video multimodal information, wherein a wide part uses a factor decomposition Machine (FM) to cross low-order features such as video categories, video duration and video quality, and a deep part performs multimodal fusion on high-order features such as titles, labels and covers of videos.

The network structure diagram of the overall algorithm in the embodiment of the present application is shown in fig. 10, and the network input features include the following two types: inputting low-order characteristics of the wide network, wherein the low-order characteristics comprise the category of a video, the video duration and the quality grade of the video; in addition, three characteristics of a title, a label and a cover page of the deep network are input. The wide & deep classification network model obtains respective prediction scores of the two paths by using the 6 video characteristics and respectively carrying out wide & deep two-path transformation, finally sums the two paths of prediction results, and transforms the prediction results to be in a range of [0, 1] by using a sigmoid nonlinear activation function so as to obtain the final popularity score.

In the model training phase of the wide & deep classification network model, the cost of misclassification can be measured using the cross entropy loss function provided by the following equation (1-1):

wherein L represents the loss result of the cross entropy loss function; avg represents the mean value; y is_nThe predicted probability for the nth sample; y is_nE {0,1} is the true tendency result for the nth sample, where 0 represents unpopular and 1 represents popular; in the model prediction stage, the prediction probability y output by the network is used, if y is larger than 0.5, the video is judged to be popular, otherwise, the video is judged to be unpopular.

Based on the network structure diagram provided in fig. 10, wherein the left half part in fig. 10 represents the wide network structure, the input features include three parts, i.e., video category, video duration, and video quality level. The main purpose of using the wide network is to introduce the memory capability of a linear model to amplify the influence of a specific characteristic combination in the tendency recognition task on the final judgment. For example, { entertainment-the eight diagrams, high quality, moderate duration } this type of video, is popular with the female community with a high probability. Specifically, in the embodiment of the present application, the embedding layer 1101 is adopted to obtain an embedding vector of an input feature, the FM layer 1002 is used as a feature crossing network, and different preprocessing methods are respectively used for three input features.

In the embodiment of the present application, for a video category feature: the video can be classified into a certain category according to the video content, and specifically classified into a first-level classification and a second-level classification, for example, the total number of the first-level classification is 42, the total number of the second-level classification is 296, and the first-level classification and the second-level classification are separately subjected to one hot coding (one hot), so that two category feature vectors x can be obtained_cls1∈R⁴²，x_cls2∈R²⁹⁶；

For the video duration feature: the video time is a continuous characteristic, so that equal-frequency binning discretization can be carried out, wherein the discretization is divided into 1-5 levels, and one hot coding is used for representing x_t∈R⁵；

For video quality features: the video quality can be scored according to prior standards such as video definition, video content and the like, the scoring result is directly used in the embodiment of the application and is graded to be 1-5, and the o ne hot code is used as x_level∈R⁵。

Then, the above x is added_cls1、x_cls2、x_tAnd x_levelFour discrete characteristics are input into the FM layer, and the output y of the wide part can be obtained through the following formula (1-2)_wide。

y_wide＝FM(x_cls1,x_cls2,x_t,x_level) (1-2)。

Please continue to refer to the network structure diagram provided in fig. 10, wherein the right half of fig. 10 shows a deep network structure, the input features used in this part are divided into two modalities, i.e., a text mode and an image mode, the text mode includes a title and a tag (tag) of a video, and the image mode is a cover of the video. Since text and image information are cross-modal, the embodiment of the present application further introduces inter-modal feature fusion based on attention mechanism (attention) to better express video content features.

First, in the feature extraction, header information is subjected toPerforming word segmentation, obtaining a word embedding vector (embedding) of each word by using word2vec, and averaging the embedding results of all words to obtain a coding vector x of the title sentence level_title(ii) a The label information is a word sequence formed by key descriptive words in the video, so that a word segmentation process is omitted, the average of all words is solved after the embedding coding result of each word is directly obtained, and the coding vector x of the label is obtained_tag(ii) a For the cover map, in the embodiment of the present application, a resnet50 network (a deep residual error network, which is a CNN-based picture feature extraction network) is directly used for feature extraction to obtain a cover map vector x_img。

After all the input features are extracted, using an attention mechanism to perform inter-modal feature fusion. Specifically, alpha is introduced into the network_title、α_tag、α_imgThree attention weights, for x respectively_title、x_tag、x_imgGiving different weights, after the weighted results are respectively subjected to linear transformation through three paths of Fully Connected networks (FC, full Connected), performing concat splicing as a feature fusion result, and performing subsequent operation, wherein the following operation is a calculation process of concat splicing treatment as shown in a formula (1-3), wherein f_i(x) Wx + B is a linear transformation result of the FC layer, where x denotes an input feature vector of the FC layer, W and B are weight parameters to be learned by the model, and the FC layer denotes a linear transformation performed on the input feature vector.

X＝concat(f₁(α_title·x_title),f₂(α_tag·x_tag),f₃(α_img·x_img)) (1-3)；

Wherein X represents the result after concat splicing treatment.

After the multi-modal fusion feature X is obtained, the output dimensionality is finally reduced to 1 dimension through 3-layer NN transformation, and the popularity score y predicted by the deep network is obtained_deep＝DNN(X)。

And finally, integrating the output results of the wide network and the deep network through the following formulas (1-4) to obtain the overall popularity score y.

y＝sigmoid(y_wide+y_deep) (1-4)。

In the embodiment of the application, the consumption logs of the specific user group are used as the marking data for judging whether the content is popular, the crowd tendency estimation problem is modeled into a two-classification problem for judging whether the content is popular with the specific crowd, the user behavior data is directly used for constructing the label, and a large amount of marking cost is saved. By using the wide & deep network model structure based on multi-mode information, the complex diversity of characteristics under the pre-estimation scene is fully considered, so that the model can more accurately depict the popularity of the content on the specific crowd. By utilizing the crowd tendency estimation result of the content, the recommending side can better solve the problem of cold start of a new user and improve the retention and the duration of the user; in addition, on the aspect of optimizing the recommendation effect for a specific crowd, the result can provide a content pool popular with the crowd, and is very important for improving the consumption index on the specific crowd.

Continuing with the exemplary structure of the information recommendation device 354 implemented as a software module provided in the embodiments of the present application, in some embodiments, as shown in fig. 2, the software module stored in the information recommendation device 354 of the memory 350 may be an information recommendation device in the server 300, including:

an obtaining module 3541, configured to obtain at least one first-class feature and at least one second-class feature of each piece of information to be recommended in an information set to be recommended; a feature cross processing module 3542, configured to perform feature cross processing on the at least one first type feature to obtain a first prediction score; a feature fusion processing module 3543, configured to perform feature fusion processing on the at least one second type feature to obtain a second prediction score; a transformation processing module 3544, configured to transform the prediction result of the first prediction score and the second prediction score to obtain a tendency score of the information to be recommended; a recommending module 3545, configured to recommend at least one piece of information to be recommended in the set of information to be recommended to a target object according to the tendency score of each piece of information to be recommended.

In some embodiments, the apparatus further comprises: the determining module is used for determining the attribute information of the target object; the set acquisition module is used for acquiring an object set formed by objects with the attribute information and a recommendation information set formed by recommendation information with interactive information between the objects in the object set; the interactive data acquisition module is used for acquiring interactive data between each object in the object set and the recommendation information for each recommendation information in the recommendation information set; the aggregation module is used for aggregating the interaction data to obtain the interaction probability of each recommendation information for the objects in the object set; and the selection module is used for selecting at least one piece of recommendation information from the recommendation information set as information to be recommended according to the interaction probability so as to form the information set to be recommended.

In some embodiments, the interaction data at least comprises exposure times, click times and browsing duration of the recommendation information; the aggregation module is further to: aggregating the exposure times, the click times and the browsing duration to obtain an average click rate and an average browsing duration for each piece of recommended information; and determining any one of the average click rate and the average browsing duration as the interaction probability of the recommendation information.

In some embodiments, the apparatus further comprises: a total average click rate determining module, configured to determine, according to the average click rate of each piece of recommendation information in the recommendation information set, a total average click rate corresponding to all objects in the object set; a total average browsing duration module, configured to determine, according to the average browsing duration of each piece of recommended information in the recommended information set, a total average browsing duration corresponding to all objects in the object set; the selection module is further configured to: selecting the recommendation information with the average click rate larger than the total average click rate from the recommendation information set as the information to be recommended; or selecting the recommendation information with the average browsing time length larger than the total average browsing time length from the recommendation information set as the information to be recommended.

In some embodiments, the first type of feature comprises at least one of an information category feature, an information length feature, and an information quality feature; the feature intersection processing module is further configured to: respectively coding the first class characteristic and the second class characteristic in the information class characteristic to correspondingly obtain a first class characteristic vector and a second class characteristic vector; carrying out discretization processing and coding processing on the information length characteristics in sequence to obtain information length characteristic vectors; coding the information quality characteristics to obtain information quality characteristic vectors; and performing the cross processing on at least one of the primary category characteristic vector, the secondary category characteristic vector, the information length characteristic vector and the information quality characteristic vector by adopting a factor decomposition machine to obtain the first prediction score.

In some embodiments, the feature intersection processing module is further to: acquiring a pre-scoring value aiming at the information to be recommended; and carrying out the coding processing on the pre-scoring value to obtain the information quality characteristic vector.

In some embodiments, the second type of feature comprises at least a text feature and an image feature; the feature fusion processing module is further configured to: determining at least one word according to the text features; acquiring a word embedding vector of each word; determining the average value of the word embedding vectors as the encoding vector of the text feature; extracting image features of the image features to obtain image feature vectors; and performing the feature fusion processing on the coding vector and the image feature vector based on an attention mechanism to obtain the second prediction score.

In some embodiments, the feature fusion processing module is further configured to: respectively performing linear transformation processing on the coding vector and the image characteristic vector based on the attention mechanism to correspondingly obtain a coding vector after linear transformation and an image characteristic vector after linear transformation; splicing the coding vector after the linear transformation and the image characteristic vector after the linear transformation to obtain a spliced vector; and activating the splicing vector to reduce the dimension of the splicing vector to one dimension to obtain the second prediction score.

In some embodiments, the feature fusion processing module is further configured to: determining a first attention weight of the encoded vector and a second attention weight of the image feature vector; weighting the encoding vector and the image feature vector with the first attention weight and the second attention weight, respectively; and respectively carrying out linear transformation processing on the weighted coding vector and the weighted image characteristic vector to correspondingly obtain the linearly transformed coding vector and the linearly transformed image characteristic vector.

In some embodiments, the transform processing module is further to: summing the first prediction score and the second prediction score to obtain a total prediction score; and performing prediction result transformation processing on the total prediction score by adopting a nonlinear activation function to obtain the tendency score of the information to be recommended.

In some embodiments, the apparatus further comprises: the processing module is used for determining the tendency score of the information to be recommended by adopting a tendency identification network; wherein the tendency recognition network is trained by: inputting a first type of sample characteristics of sample information and a second type of sample characteristics of the sample information into the tendency identification network; performing feature cross processing on the first type of sample features through a feature cross network of the tendency identification network to obtain a first sample prediction score; performing feature fusion processing on the second type sample features through a feature fusion network of the tendency identification network to obtain a second sample prediction score; performing prediction result transformation processing on the first sample prediction score and the second sample prediction score through a transformation processing layer of the tendency identification network to obtain a sample tendency score of the sample information; inputting the sample tendency score into a preset loss model to obtain a loss result; and according to the loss result, correcting parameters in the feature cross network and the feature fusion network to obtain a trained tendency recognition network.

It should be noted that the description of the apparatus in the embodiment of the present application is similar to the description of the method embodiment, and has similar beneficial effects to the method embodiment, and therefore, the description is not repeated. For technical details not disclosed in the embodiments of the apparatus, reference is made to the description of the embodiments of the method of the present application for understanding.

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 method of the embodiment of the present application.

Embodiments of the present application provide a storage medium having stored therein executable instructions, which when executed by a processor, will cause the processor to perform a method provided by embodiments of the present application, for example, the method as illustrated in fig. 3.

In some embodiments, the storage medium may be a computer-readable storage medium, such as a Ferroelectric Random Access Memory (FRAM), a Read Only Memory (ROM), a Programmable Read Only Memory (PROM), an Erasable Programmable Read Only Memory (EPROM), a charged Erasable Programmable Read Only Memory (EEPROM), a flash Memory, a magnetic surface Memory, an optical disc, or a Compact disc Read Only Memory (CD-ROM), among other memories; or may be various devices including one or any combination of the above memories.

In some embodiments, executable instructions may be written in any form of programming language (including compiled or interpreted languages), in the form of programs, software modules, scripts or code, and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

By way of example, executable instructions may correspond, but do not necessarily have to correspond, to files in a file system, and may be stored in a portion of a file that holds other programs or data, such as in one or more scripts in a hypertext 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). By way of example, executable instructions may be deployed to be executed on one computing device or on multiple computing devices at one site or distributed across multiple sites and interconnected by a communication network.

The above description is only an example of the present application, and is not intended to limit the scope of the present application. Any modification, equivalent replacement, and improvement made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. An information recommendation method, comprising:

acquiring at least one first type feature and at least one second type feature of each piece of information to be recommended in an information set to be recommended;

performing feature cross processing on the at least one first type feature to obtain a first prediction score;

performing feature fusion processing on the at least one second type feature to obtain a second prediction score;

performing prediction result transformation processing on the first prediction score and the second prediction score to obtain a tendency score of the information to be recommended;

and recommending at least one piece of information to be recommended in the information set to be recommended to a target object according to the tendency score of each piece of information to be recommended.

2. The method of claim 1, further comprising:

determining attribute information of the target object;

acquiring an object set formed by objects with the attribute information and a recommendation information set formed by recommendation information with interactive information between the objects in the object set;

for each piece of recommendation information in the recommendation information set, acquiring interaction data between each object in the object set and the recommendation information;

aggregating the interaction data to obtain interaction probability of each recommendation information for the objects in the object set;

and selecting at least one piece of recommendation information from the recommendation information set as information to be recommended according to the interaction probability to form the information set to be recommended.

3. The method of claim 2, wherein the interaction data comprises at least exposure times, click times and browsing duration of the recommended information;

the aggregating the interaction data to obtain the interaction probability of each recommendation information for the objects in the object set includes:

aggregating the exposure times, the click times and the browsing duration to obtain an average click rate and an average browsing duration for each piece of recommended information;

and determining any one of the average click rate and the average browsing duration as the interaction probability of the recommendation information.

4. The method of claim 3, further comprising:

determining a total average click rate corresponding to all objects in the object set according to the average click rate of each piece of recommendation information in the recommendation information set;

determining a total average browsing duration corresponding to all objects in the object set according to the average browsing duration of each piece of recommended information in the recommended information set;

the selecting at least one piece of recommendation information from the recommendation information set as information to be recommended according to the interaction probability includes:

selecting the recommendation information with the average click rate larger than the total average click rate from the recommendation information set as the information to be recommended; alternatively, the first and second electrodes may be,

and selecting the recommendation information with the average browsing time length larger than the total average browsing time length from the recommendation information set as the information to be recommended.

5. The method of claim 1, wherein the first type of feature comprises at least one of an information category feature, an information length feature, and an information quality feature;

the performing feature cross processing on the at least one first type feature to obtain a first prediction score includes:

respectively coding the first class characteristic and the second class characteristic in the information class characteristic to correspondingly obtain a first class characteristic vector and a second class characteristic vector;

carrying out discretization processing and coding processing on the information length characteristics in sequence to obtain information length characteristic vectors;

coding the information quality characteristics to obtain information quality characteristic vectors;

and performing the cross processing on at least one of the primary category characteristic vector, the secondary category characteristic vector, the information length characteristic vector and the information quality characteristic vector by adopting a factor decomposition machine to obtain the first prediction score.

6. The method of claim 5, wherein the encoding the information quality characteristic to obtain an information quality characteristic vector comprises:

acquiring a pre-scoring value aiming at the information to be recommended;

and carrying out the coding processing on the pre-scoring value to obtain the information quality characteristic vector.

7. The method of claim 1, wherein the second type of features comprises at least text features and image features;

the performing feature fusion processing on the at least one second type feature to obtain a second prediction score includes:

determining at least one word according to the text features;

acquiring a word embedding vector of each word;

determining the average value of the word embedding vectors as the encoding vector of the text feature;

extracting image features of the image features to obtain image feature vectors;

and performing the feature fusion processing on the coding vector and the image feature vector based on an attention mechanism to obtain the second prediction score.

8. The method according to claim 7, wherein the performing the feature fusion process on the encoded vector and the image feature vector based on the attention mechanism to obtain the second prediction score comprises:

respectively performing linear transformation processing on the coding vector and the image characteristic vector based on the attention mechanism to correspondingly obtain a coding vector after linear transformation and an image characteristic vector after linear transformation;

splicing the coding vector after the linear transformation and the image characteristic vector after the linear transformation to obtain a spliced vector;

and activating the splicing vector to reduce the dimension of the splicing vector to one dimension to obtain the second prediction score.

9. The method according to claim 8, wherein the performing linear transformation processing on the encoded vector and the image feature vector based on the attention mechanism to obtain a linearly transformed encoded vector and a linearly transformed image feature vector respectively comprises:

determining a first attention weight of the encoded vector and a second attention weight of the image feature vector;

weighting the encoding vector and the image feature vector with the first attention weight and the second attention weight, respectively;

and respectively carrying out linear transformation processing on the weighted coding vector and the weighted image characteristic vector to correspondingly obtain the linearly transformed coding vector and the linearly transformed image characteristic vector.

10. The method according to claim 1, wherein the transforming the prediction results of the first prediction score and the second prediction score to obtain the tendency score of the information to be recommended comprises:

summing the first prediction score and the second prediction score to obtain a total prediction score;

and performing prediction result transformation processing on the total prediction score by adopting a nonlinear activation function to obtain the tendency score of the information to be recommended.

11. The method according to any one of claims 1 to 10, further comprising: determining the tendency score of the information to be recommended by adopting a tendency identification network;

wherein the tendency recognition network is trained by:

inputting a first type of sample characteristics of sample information and a second type of sample characteristics of the sample information into the tendency identification network;

performing feature cross processing on the first type of sample features through a feature cross network of the tendency identification network to obtain a first sample prediction score;

performing feature fusion processing on the second type sample features through a feature fusion network of the tendency identification network to obtain a second sample prediction score;

performing prediction result transformation processing on the first sample prediction score and the second sample prediction score through a transformation processing layer of the tendency identification network to obtain a sample tendency score of the sample information;

inputting the sample tendency score into a preset loss model to obtain a loss result;

and according to the loss result, correcting parameters in the feature cross network and the feature fusion network to obtain a trained tendency recognition network.

12. An information recommendation apparatus, comprising:

the acquisition module is used for acquiring at least one first type feature and at least one second type feature of each piece of information to be recommended in the information set to be recommended;

the characteristic cross processing module is used for carrying out characteristic cross processing on the at least one first type characteristic to obtain a first prediction score;

the characteristic fusion processing module is used for carrying out characteristic fusion processing on the at least one second type characteristic to obtain a second prediction score;

the transformation processing module is used for carrying out transformation processing on the prediction results of the first prediction score and the second prediction score to obtain the tendency score of the information to be recommended;

and the recommending module is used for recommending at least one piece of information to be recommended in the information set to be recommended to a target object according to the tendency score of each piece of information to be recommended.

13. An information recommendation apparatus characterized by comprising:

a memory for storing executable instructions; a processor for implementing the information recommendation method of any one of claims 1 to 11 when executing the executable instructions stored in the memory.

14. A computer-readable storage medium having stored thereon executable instructions for causing a processor to execute the executable instructions to implement the information recommendation method of any one of claims 1 to 11.