Material recommendation method, device and equipment and computer storage medium

Through the multi-type behavior interest model, user behavior is uniformly processed, and shared characteristics and interest expression characteristics are generated, which solves the problem of inconsistent user interest expression in the existing recommendation system, and achieves more accurate material recommendations and reduces data complexity.

CN118972436APending Publication Date: 2024-11-15MICRO DREAM TECHTRONIC NETWORK TECH CHINACO
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410984973.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing recommendation system cannot accurately predict users' interest preferences for materials, because the multiple behaviors performed by the same material by users are not uniformly processed, resulting in inconsistent expression of interest, and ignore the dependence and superimposed strengthening relationships between different user behavior types. Accurate recommendations under different task objectives.

Method used

Through a multi-type behavior interest model, combining user characteristics and material characteristics, a multi-head target attention mechanism and a multi-layer perceptron network are adopted to uniformly process multiple user behaviors, generate shared features and interest expression features, and realize accurate representation of user interests and materials Accurate recommendations.

Benefits of technology

It realizes more accurate user interest characterization and material recommendation, reduces the complexity of data processing and modeling, and can better express the user's real-time interest changes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118972436A_ABST
    Figure CN118972436A_ABST
Patent Text Reader

Abstract

The invention provides a material recommendation method, device and equipment and a computer storage medium, and the method comprises the steps: obtaining material features of candidate materials, user features of a target user and multi-type behavior sequence features based on a recommendation request sent by the target user; obtaining shared features based on the user features and the material features through a sharing module of the multi-type behavior interest model; determining a multi-type behavior interest expression feature corresponding to each task target based on the material features and the multi-type behavior sequence features through a multi-target attention mechanism module of a multi-type behavior interest model; through a multi-layer perceptron network of a multi-type behavior interest model, based on each multi-type behavior interest expression feature and sharing feature, obtaining a prediction result of the target user on the candidate material for the corresponding task target; and based on the prediction result, performing material recommendation to the target user. According to the method and the device, the interest preference of the user on the material can be more accurately represented.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a material recommendation method, device, equipment and computer storage medium. Background Art

[0002] In the information flow recommendation system, users will generate a series of feedback behaviors in response to the recommendation results of pictures, texts, and videos given by the recommendation system, such as playing, liking, forwarding, commenting, clicking on pictures, and complaining. These user behaviors are driven by users' long-term and short-term interests. At the same time, for specific user behaviors, the degree of user interest reflected is also different. For example, the longer the user reads the text, text, and video blog posts, the more the user likes them; the degree of preference for blog posts behind the user's various user behaviors such as clicking on pictures, clicking on texts, liking, forwarding, and commenting in the blog posts is also different.

[0003] Most existing recommendation systems use the method of splitting multiple types of user behavior sequences for modeling. Multiple user behaviors performed by the same user on the same item are not processed uniformly, resulting in the same item being expressed as different interests, which makes it impossible for the recommendation system to accurately predict the user's interest preference for the item, and thus unable to achieve accurate recommendations. Summary of the invention

[0004] The present application provides a material recommendation method, device, equipment and computer storage medium to more accurately characterize the user's interest preference for materials.

[0005] The technical solution of this application is implemented as follows:

[0006] The present application provides a material recommendation method, which includes: based on a recommendation request sent by a target user, obtaining material features of a candidate material, user features of a target user, and multi-type behavior sequence features; the multi-type behavior sequence features are used to characterize the target user's various types of user behaviors on multiple exposed materials within a set time range; through a sharing module of a multi-type behavior interest model, based on user features and material features, a shared feature is obtained; the shared feature is a fusion feature of the user feature and the material feature; through a multi-head target attention mechanism module of a multi-type behavior interest model, based on material features and multi-type behavior sequence features, a multi-type behavior interest expression feature corresponding to each task target is determined; the multi-type behavior interest expression feature is used to characterize the target user's interest preference for the material under the corresponding task target; through a multi-layer perceptron network of a multi-type behavior interest model, based on each multi-type behavior interest expression feature and shared feature, a prediction result of the target user for the candidate material for the corresponding task target is obtained; based on the prediction result, a material is recommended to the target user.

[0007] In some possible implementations, before obtaining material features of candidate materials, user features of target users, and multi-type behavior sequence features, the method further includes: in response to a recommendation request sent by the target user, obtaining material information of the candidate materials, user information of the target user, and historical behavior marks of the target user within a set time range; performing feature extraction processing on the material information to obtain material features; performing feature extraction processing on the user information to obtain user features; performing feature extraction processing on the historical behavior marks to obtain multi-type behavior sequence features.

[0008] In some possible implementations, before obtaining the historical behavior marks of the target user within a set time range, the method also includes: collecting the historical behavior information of the target user on multiple exposed materials within the set time range in the target user's behavior log; encoding the historical behavior information to obtain a binary array having a mapping relationship with the historical behavior information; performing integer conversion on the binary array to obtain historical behavior marks having a mapping relationship with the binary array, and storing the historical behavior marks in a graph storage database.

[0009] In some possible implementations, the material information includes first material information and second material information, the first material information is used to characterize the material code of the candidate material, and the second material information is used to characterize the material content of the candidate material; performing feature extraction processing on the material information to obtain material characteristics, including: performing feature extraction processing on the first material information to obtain a first material characteristic; performing feature extraction processing on the second material information to obtain a second material characteristic; and combining the first material characteristic and the second material characteristic into a material characteristic.

[0010] In some possible implementations, a multi-type behavior interest expression feature corresponding to each task target is determined based on material features and multi-type behavior sequence features through a multi-head target attention mechanism module of a multi-type behavior interest model, including: extracting sub-multi-type behavior interest expression features corresponding to each task target in the multi-type behavior interest expression features through the multi-head target attention mechanism module; performing function mapping processing on each sub-multi-type behavior interest expression feature and the first material feature to obtain the multi-type behavior interest expression feature corresponding to each task target.

[0011] In some possible implementations, shared features are obtained based on user features and material features through a sharing module of multi-type behavioral interest models, including: fusing user features and second material features through a sharing module to obtain shared features.

[0012] In some possible implementations, material recommendations are made to target users based on the prediction results, including: sorting candidate materials based on their prediction results under each task objective; and recommending target materials to the target user from the candidate materials based on the sorting results.

[0013] In some possible implementations, after recommending materials to the target user based on the prediction results, the method also includes: collecting multiple types of user behaviors of the target user on the newly exposed target material to obtain real-time user behavior; encoding and integer converting the real-time user behavior to obtain the real-time mark of the target user, and storing the real-time behavior mark in a graph storage database; generating training samples based on the material information of the target material, the user information of the target user and the real-time behavior mark, using the training samples to perform online training on a multi-type behavior interest model, adjusting the model parameters of the multi-type behavior interest model, and obtaining an updated multi-type behavior interest model.

[0014] The present application provides a material recommendation device, which includes: a data acquisition module, which is used to acquire material features of candidate materials, user features of target users and multi-type behavior sequence features based on a recommendation request sent by a target user; the multi-type behavior sequence features are used to characterize the target user's various types of user behaviors on multiple exposed materials within a set time range; a first data processing module, which is used to obtain shared features based on user features and material features through a sharing module of a multi-type behavior interest model; the shared features are fusion features of user features and material features; a second data processing module, which is used to determine the multi-type behavior interest expression features corresponding to each task target based on material features and multi-type behavior sequence features through a multi-head target attention mechanism module of a multi-type behavior interest model; the multi-type behavior interest expression features are used to characterize the target user's interest preference for materials under the corresponding task target; a prediction module, which is used to obtain the target user's prediction results for candidate materials for the corresponding task target based on each multi-type behavior interest expression feature and shared features through a multi-layer perceptron network of a multi-type behavior interest model; and a recommendation module, which is used to recommend materials to the target user based on the prediction results.

[0015] In some possible implementations, the above-mentioned device also includes: a feature extraction module, which is used to respond to a recommendation request sent by a target user, obtain material information of a candidate material, user information of a target user, and historical behavior marks of a target user within a set time range; perform feature extraction processing on the material information to obtain material features; perform feature extraction processing on the user information to obtain user features; perform feature extraction processing on the historical behavior marks to obtain multi-type behavior sequence features.

[0016] In some possible implementations, the above-mentioned device also includes: a behavior tag generation module, which is used to collect the historical behavior information of the target user on multiple exposed materials within a set time range in the target user's behavior log; encode the historical behavior information to obtain a binary array having a mapping relationship with the historical behavior information; perform integer conversion processing on the binary array to obtain a historical behavior tag having a mapping relationship with the binary array, and store the historical behavior tag in a graph storage database.

[0017] In some possible implementations, the material information includes first material information and second material information, the first material information is used to characterize the material code of the candidate material, and the second material information is used to characterize the material content of the candidate material; a feature extraction module is used to perform feature extraction processing on the first material information to obtain a first material feature; perform feature extraction processing on the second material information to obtain a second material feature; and combine the first material feature and the second material feature into a material feature.

[0018] In some possible implementations, the second data processing module is used to extract sub-multi-type behavior interest expression features corresponding to each task target from the multi-type behavior interest expression features through a multi-head target attention mechanism module; perform function mapping processing on each sub-multi-type behavior interest expression feature and the first material feature to obtain the multi-type behavior interest expression features corresponding to each task target.

[0019] In some possible implementations, the first data processing module is used to fuse the user features and the second material features through the sharing module to obtain the shared features.

[0020] In some possible implementations, the recommendation module is used to sort the candidate materials based on the prediction results of the candidate materials under each task goal; and recommend the target material to the target user from the candidate materials based on the sorting results.

[0021] In some possible implementations, the above-mentioned device also includes: a training module, which is used to collect various types of user behaviors of the target user on the newly exposed target material to obtain real-time user behavior; encode the real-time user behavior and perform integer conversion processing to obtain the real-time behavior mark of the target user, and store the real-time behavior mark in a graph storage database; generate training samples based on the material information of the target material, the user information of the target user and the real-time behavior mark, use the training samples to perform online training on the multi-type behavior interest model, adjust the model parameters of the multi-type behavior interest model, and obtain an updated multi-type behavior interest model.

[0022] The present application provides an electronic device, comprising: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided by the present application.

[0023] The present application provides a computer storage medium storing executable instructions for implementing the method provided by the present application when the executable instructions are executed by a processor.

[0024] The present application provides a computer program product, including a computer program or instructions, for implementing the method provided by the present application when the computer program or instructions are executed by a processor.

[0025] This application has the following beneficial effects:

[0026] In this application, through the multi-head target attention mechanism module of the multi-type behavior interest model, based on the material characteristics and multi-type behavior sequence characteristics, the multi-type behavior interest expression characteristics corresponding to each task goal are determined, and various types of user behaviors can be processed in a unified manner, thereby more accurately representing the target user's interest preference for the material.

[0027] Furthermore, based on various types of user behaviors, it is predicted that the target user will perform various types of user behaviors on the candidate materials, so that the prediction results are more in line with the user's interest preferences for the materials, and more accurate material recommendations are achieved.

[0028] Furthermore, by introducing encoding processing and integer conversion processing to obtain historical behavior tags, it is possible to encode multiple types of user behaviors with one tag, reduce the complexity of separate storage of multiple types of user behavior data, and reduce the complexity of data processing and modeling in the material recommendation process.

[0029] Furthermore, the encoded historical behavior markers are uniformly embedded into vectors (i.e., feature extraction processing), which can model the superposition and reinforcement relationship between different types of user behaviors and the behavioral dependency relationship for different task objectives.

[0030] Furthermore, based on the real-time user behavior of target users towards newly exposed target materials, online training of multi-type behavioral interest models can better express the real-time interest changes of target users. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the architecture of the recommendation system provided in the embodiment of the present application;

[0032] Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0033] Figure 3 This is an optional flow chart of the material recommendation method provided in the embodiment of the present application;

[0034] Figure 4It is an optional structural diagram of a multi-type behavior interest model provided in an embodiment of the present application;

[0035] Figure 5 It is an optional flow chart of the material recommendation method provided in the embodiment of the present application. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0037] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it can be 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.

[0038] If similar descriptions of "first / second" appear in the application documents, the following explanation is added. In the following description, the terms "first / second / third" involved are merely to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first / second / third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0040] In the information flow recommendation system, users will generate a series of feedback behaviors in response to the recommendation results of pictures, texts, and videos given by the recommendation system, such as playing, liking, forwarding, commenting, clicking on pictures, and complaining. These user behaviors are driven by the long-term and short-term interests of users. At the same time, for specific user behaviors, the degree of user interest reflected is also different. For example, the longer the user reads the text, text, and video blog posts, the more the user likes them; the degree of preference for blog posts behind the user's clicking on pictures, texts, liking, forwarding, and commenting in the blog posts is also different. Therefore, hybrid modeling of all the user's recent feedback behaviors on the recommendation system can characterize the user's interests in a fine-grained manner, thereby achieving accurate recommendations.

[0041] Existing recommendation systems mostly use split sequence type modeling to model user sequence interests with multiple user behavior types. For example, in the DIN (deep interest network) model in the A e-commerce advertising scenario, different types of sequence data are constructed for user behavior sequences under different user behavior types, such as user click sequences, user order sequences, etc., and then different sequences are modeled separately. In the DFN (discrete fracture network) model in the B application's "Look at the Information Flow" scenario, the user's click sequence, implicit negative feedback sequence (behavior of being exposed but not clicked), and explicit negative feedback sequence (behavior of clicking the "not interested" button) are similarly distinguished, and then they are modeled separately and then merged to characterize user interests. In the DMT (deep multifaceted transformers) model in the B e-commerce recommendation scenario, the user behavior sequence is also split into: click sequence, shopping cart sequence, and purchase sequence to model user interests at different granularities.

[0042] In existing recommendation tasks, most of the user behavior sequences of various types are modeled by splitting the user behavior sequences. The main problems are: 1) ignoring the situation where there are multiple user behaviors in an item at the same time, such as when a user clicks, adds to cart, and purchases a certain product. In this case, the user's interest in the product is split into three sequences and is not processed uniformly. At this time, the same product is expressed as different interests in different sequences; 2) ignoring the dependency and superposition reinforcement relationship between different user behavior types. For example, when a user reads for 3 seconds after clicking a blog post and reads for 10 seconds after clicking a blog post, clicks and reading are positive superposition relationships, and the reading time will strengthen the interest intensity implied by the user's click behavior; and when a user clicks on a picture to view it, and then clicks to complain after not liking the picture, the behavior at this time is a negative superposition relationship, and the complaint should weaken the interest intensity behind the click behavior; 3) ignoring the different dependency strengths of different user behavior types under different task goals. For example, when modeling the user reading time goal, the correlation between the reading time of the sequence and the goal will be stronger than forwarding, commenting, and liking behaviors. Based on this, the recommendation system cannot accurately predict users' interest preferences in materials, and thus cannot achieve accurate recommendations.

[0043] In response to the above problems, an embodiment of the present application provides a material recommendation method to more accurately characterize the user's interest preference in materials.

[0044] The following describes an exemplary application of the electronic device provided by the embodiment of the present application. The electronic device provided by the embodiment of the present application can be implemented as various types of user terminals such as a laptop computer, a tablet computer, a desktop computer, a mobile device (for example, a mobile phone, a wearable smart watch, a dedicated messaging device), and can also be implemented as a server. The following describes an exemplary application when the electronic device is implemented as a server.

[0045] See also Figure 1 , Figure 1 It is a schematic diagram of the architecture of the recommendation system provided in an embodiment of the present application. The terminals (terminal 40-1 and terminal 40-2 are shown as examples) are connected to the server 200 via the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0046] In some possible implementations, user A can publish material A through terminal 40-1, and user B can publish material B through terminal 40-2. Material A and material B are uploaded to server 200 through network 300. Server 200 can store material A and material B in database 500. In order to make accurate material recommendations according to the interests of users, server 200 can push the materials that match the interests of users most to users through sorting processing, and the materials pushed to users can be displayed in the graphical interface of the terminal (graphical interface 41-1 and graphical interface 41-2 are shown as examples).

[0047] In some embodiments, the server 200 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, and big data and artificial intelligence platforms. The terminal may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart watch, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present invention.

[0048] See also Figure 2 , Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, Figure 2The electronic device 400 shown may be the above-mentioned terminal and / or server 200. The electronic device 400 includes: at least one processor 410, a memory 450, at least one network interface 420 and a user interface 430. The various components in the electronic device 400 are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 440 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 440 is not shown in FIG. Figure 2 Various buses are labeled as bus system 440 .

[0049] Processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0050] The user interface 430 includes one or more output devices 431 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0051] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0052] The memory 450 includes a volatile memory or a non-volatile memory, and may also include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0053] In some embodiments, memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0054] Operating system 451, 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;

[0055] A network communication module 452, for reaching other computing devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 include: Bluetooth, wireless compatibility certification (WiFi), and universal serial bus (USB), etc.;

[0056] a presentation module 453 for enabling presentation of information via one or more output devices 431 (e.g., display screen, speaker, etc.) associated with the user interface 430 (e.g., a user interface for operating peripherals and displaying content and information);

[0057] The input processing module 454 is used to detect one or more user inputs or interactions from one of the one or more input devices 432 and translate the detected inputs or interactions.

[0058] In some embodiments, the material recommendation device provided in the embodiments of the present application can be implemented in a software manner. Figure 4 A material recommendation device 455 stored in a memory 450 is shown, which may be software in the form of a program or plug-in, etc. The material recommendation device 455 includes the following software modules: a data acquisition module 4551, a first data processing module 4552, a second data processing module 4553, a prediction module 4554 and a recommendation module 4555.

[0059] These modules are logical, so they can be arbitrarily combined or further divided according to the functions to be implemented. The functions of each module will be explained below.

[0060] In other embodiments, the material recommendation device provided in the embodiments of the present application can be implemented in hardware. As an example, the device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the material recommendation method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application specific integrated circuits (application specific integrated circuit, ASIC,), DSP, programmable logic device (programmable logic device, PLD,), complex programmable logic device (complex programmable logic device, CPLD), field programmable gate array (field-programmable gate array, FPGA) or other electronic components.

[0061] The material recommendation method provided in the embodiment of the present application will be described below in conjunction with the exemplary application and implementation of the electronic device provided in the embodiment of the present application.

[0062] It should be noted that, in the embodiment of the present application, the material recommendation system (hereinafter referred to as the recommendation system) is used to recommend materials to target users, and the materials may be blog posts published by other users in the material recommendation system.

[0063] See also Figure 3 , Figure 3 This is an optional flow chart of the material recommendation method provided in the embodiment of the present application. Figure 3 The steps shown illustrate the material recommendation method.

[0064] S301, based on the recommendation request sent by the target user, obtain material features of the candidate material, user features of the target user, and multi-type behavior sequence features.

[0065] Among them, the multi-type behavior sequence features are used to characterize the target user's various types of user behaviors on multiple exposed materials within a set time range.

[0066] It is understandable that the target user can generate multiple types of user behaviors for an exposed material, such as interaction, click, and collection; the multi-type behavior sequence features can characterize the multiple exposed materials on which the target user has performed user behaviors, as well as which user behaviors have occurred on the exposed materials respectively. The target user sends a recommendation request to the recommendation system, and the recommendation system can obtain the material characteristics of the candidate material, the user characteristics of the target user, and the multi-type behavior sequence characteristics based on the recommendation request. S302, through the sharing module of the multi-type behavior interest model, based on the user characteristics and material characteristics, obtain the shared features. Among them, the shared features are the fusion features of the user characteristics and the material characteristics.

[0067] It can be understood that the multi-type behavior interest model includes a sharing module. The user features and the material features are input into the sharing module, and the sharing module can fuse the user features and the material features and output the shared features.

[0068] S303, through the multi-head target attention mechanism module of the multi-type behavior interest model, based on the material characteristics and the multi-type behavior sequence characteristics, determine the multi-type behavior interest expression characteristics corresponding to each task target.

[0069] Among them, multi-type behavioral interest expression features are used to characterize the target user's interest preference for materials under the corresponding task objectives.

[0070] It can be understood that the multi-type behavior interest model includes a multi-head target attention mechanism module. The recommendation system inputs material features and multi-type behavior sequence features into the multi-head target attention mechanism module, and the multi-head target attention mechanism module outputs multi-type behavior interest expression features. The multi-type behavior interest expression features can characterize the user's interest in the historical recommended materials according to different task goals. The task goal is a user behavior that the user can perform on the material.

[0071] In some embodiments, taking the case where the task objectives include clicks, interactions, and viewing time as an example, the multi-type behavioral interest expression features may include: multi-type behavioral interest expression features for clicks, multi-type behavioral interest expression features for interactions, and multi-type behavioral interest expression features for viewing time. It should be noted that the task objectives are set according to actual needs, and the types of task objectives are not limited to the user behaviors listed above, but can also be other user behaviors; the number of task objectives can be one or more, and accordingly, the number of multi-type behavioral interest expression features can be one or more, and the embodiments of the present application do not specifically limit this.

[0072] S304, through the multi-layer perceptron network of the multi-type behavior interest model, based on each multi-type behavior interest expression feature and shared feature, obtain the target user's prediction result for the candidate material for the corresponding task goal.

[0073] It can be understood that the multi-type behavior interest model includes a multilayer perceptron network (multilayer perceptron, MLP). The multilayer perceptron network can perform function mapping processing on the input features (i.e., each multi-type behavior interest expression feature and shared feature) through a preset function, and output the predicted value (i.e., the prediction result) of the probability of the target user performing the user behavior corresponding to the task target on the candidate material.

[0074] S305: Recommend materials to target users based on the prediction results.

[0075] It can be understood that the multi-layer perceptron network outputs the predicted value (i.e., the predicted result) of the probability of the target user performing the user behavior corresponding to the task goal for the candidate material. Multiple candidate materials can be sorted by the predicted value of the probability of the target user performing the user behavior corresponding to the task goal for different candidate materials, and then materials can be recommended to the target user based on the sorting result.

[0076] In some possible implementations, before executing the above S301, the material recommendation method may also include: in response to a recommendation request sent by a target user, obtaining material information of a candidate material, user information of a target user, and historical behavior marks of the target user within a set time range; performing feature extraction processing on the material information to obtain material features; performing feature extraction processing on the user information to obtain user features; performing feature extraction processing on historical behavior marks to obtain multi-type behavior sequence features.

[0077] Among them, the historical behavior tag is used to characterize the various types of user behaviors performed by the target user on the historical recommended materials (i.e., the exposed materials). The multi-type behavior sequence feature is used to indicate the combination of various types of user behaviors performed by the target user on the historical recommended materials. For example, if user A clicks and watches the historical recommended material A, then the combination of various types of user behaviors performed by user A on the historical recommended material A is [click, watch].

[0078] It is understandable that after the recommendation system recommends historical recommended materials to the target user, various types of user behaviors performed by the target user on the historical recommended materials can be recorded in the form of historical behavior tags, and the historical behavior tags can be stored in the user behavior graph database. User behaviors may include but are not limited to: clicks, comments, forwarding, favorites, consumption, playback, clicking on pictures, and complaints. When the target user wants to obtain the materials recommended by the recommendation system, he can send a recommendation request to the recommendation system. Based on the received recommendation request, the recommendation system obtains the user's historical behavior tags from the user behavior graph database, and obtains the material information of the candidate materials and the user information of the target user from the user and material information database. Among them, the candidate materials are materials newly released by other users in the recommendation system, and the candidate materials in the user and material information database are updated in real time.

[0079] In some embodiments, the material information of the candidate material may include: the material identification of the candidate material and the material content of the candidate material, wherein the material identification of the candidate material is the material ID (identity) of the candidate material. The user information of the target user may include: the user identification of the user.

[0080] In some embodiments, in addition to the user ID, user information may also include but is not limited to: user age, user gender, device number of the electronic device used by the user, occupation, etc. In addition to the material ID and material content, material information may also include but is not limited to: the number of words in the material (if the material includes text), the release time of the material, the current number of readings, and the exposure click rate of the material, etc.

[0081] In some embodiments, the material content of the candidate material refers to the content displayed on the display interface of the terminal when the candidate material is published in the recommendation system, for example, text, pictures, videos and other content.

[0082] In some embodiments, the architecture of the recommendation system includes an embedding layer for performing feature extraction processing on input data.

[0083] In some embodiments, after obtaining the material information of the candidate material, the user information of the target user, and the historical behavior marks of the target user within a set time range, the recommendation system performs feature extraction processing on the material information of the candidate material, the user information of the target user, and the historical behavior marks of the target user within the set time range through the embedding layer to obtain material features, user features, and multi-type behavior sequence features.

[0084] In one example, the historical behavior mark of the target user is input into the embedding layer. After the embedding layer performs feature extraction on the historical behavior mark, the output multi-type behavior sequence features can be expressed as user_behaviour_mat = [mid_seq, multiple_act_seq] ∈ R L×(h+h) , where mid_seq represents the material ID sequence of historical recommended materials, multiple_act_seq represents the sequence of combinations of multiple types of user behaviors performed by the target user on historical recommended materials; R is the real number space, and L and h both represent dimensions.

[0085] In some embodiments, the recommendation system may also obtain context information in the process of obtaining the material information of the candidate material, the user information of the target user, and the historical behavior mark of the target user within a set time range. The context information may include the current time and the current longitude and latitude. In the case of obtaining the context information, in the process of generating user features based on the user information, the context information and the user information may be combined, and feature extraction may be performed on the context information and the user information to obtain user features. In this case, the user features may represent not only the user information of the user, but also the context information.

[0086] In some possible implementations, before obtaining the historical behavior marks of the target user within a set time range, the above-mentioned material recommendation method may also include: collecting the historical behavior information of the target user on multiple exposed materials within the set time range in the target user's behavior log; encoding the historical behavior information to obtain a binary array having a mapping relationship with the historical behavior information; performing integer conversion on the binary array to obtain historical behavior marks having a mapping relationship with the binary array, and storing the historical behavior marks in a graph storage database.

[0087] It can be understood that by preprocessing the historical behavior information (ie, the above encoding process and the above integer conversion process), the historical behavior mark can be obtained.

[0088] In some embodiments, the recommendation system collects the target user's historical behavior information on multiple exposed materials within a set time range in the target user's behavior log, encodes the historical behavior information into a binary array of a preset length; and performs integer conversion on the binary data to obtain an integer tag (i.e., historical behavior tag) corresponding to the above binary data.

[0089] In one example, the recommendation system can encode the historical behavior information of user A and generate a 31-bit binary array (i.e., a 0 / 1 array), where each digit in the array corresponds to a user behavior, 0 indicates that the user did not perform the user behavior, and 1 indicates that the user performed the user behavior. Assume that the above 31-bit binary array is [1,0,0,1,0,0,0,0,0,1,0,0,0,0,...,0], which means that user A has three user behaviors on material A, including clicking on the personal page of user B who published material A, and user A's reading time ≥3 seconds and less than 8 seconds. By converting the binary data into an integer, the integer mark (i.e., the historical behavior mark) is 521, and 521 can represent the above three user behaviors.

[0090] In an embodiment of the present application, by encoding the historical behavior information to obtain a binary array, and then converting the binary array into integers to obtain historical behavior tags, the amount of data of the target user's historical behavior information can be reduced, thereby reducing the modeling complexity of multi-type behavior interest models.

[0091] In some possible implementations, the material information includes first material information and second material information, the first material information is used to characterize the material code of the candidate material, and the second material information is used to characterize the material content of the candidate material; the above-mentioned feature extraction processing of the material information to obtain the material characteristics may include: performing feature extraction processing on the first material information to obtain the first material characteristics; performing feature extraction processing on the second material information to obtain the second material characteristics; and combining the first material characteristics and the second material characteristics into a material characteristic.

[0092] It can be understood that the material information can be divided into the first material information and the second material information. In the process of feature extraction processing, the embedding layer can perform feature extraction processing on the first material information and the second material information respectively to obtain the first material feature and the second material feature. The first material feature and the second material feature constitute the material feature. Among them, the first material feature has a mapping relationship with the material ID of the candidate material, and the second material feature has a mapping relationship with the material content of the candidate material.

[0093] In one example, the material ID of the candidate material is input into the embedding layer, and the embedding layer performs feature extraction processing on the material ID of the candidate material. The output first material feature can be expressed as item_feat=[mid]∈R 1×h , where mid represents the material ID of the candidate material and h represents the dimension.

[0094] In some possible implementations, the above S303 may include: extracting sub-multi-type behavior interest expression features corresponding to each task target from the multi-type behavior interest expression features through a multi-head target attention mechanism module; performing function mapping processing on each sub-multi-type behavior interest expression feature and the first material feature to obtain the multi-type behavior interest expression features corresponding to each task target.

[0095] It can be understood that user behaviors can include many types. In order to facilitate data calculation and processing, only several representative user behaviors are selected as task targets. For example, assuming that the preset representative user behaviors include: click, interaction and viewing time, then the prediction values ​​output by the multi-type behavior interest model in the recommendation system are the prediction values ​​of the probability of the target user clicking on the candidate material, the prediction values ​​of the probability of the target user interacting with the candidate material, and the prediction values ​​of the target user's viewing time of the candidate material. The multi-head target attention mechanism module is used to split the multi-type behavior interest expression feature into sub-multi-type behavior interest expression features corresponding to each task target according to the preset task target. Assuming that there are multiple task targets, the multi-type behavior interest expression feature includes multiple sub-multi-type behavior interest expression features; multiple sub-multi-type behavior interest expression features correspond to multiple task targets one by one; then the multiple sub-multi-type behavior interest expression features are respectively subjected to function mapping processing with the first material feature, so as to obtain multiple multi-type behavior interest expression features, and each task target corresponds to a multi-type behavior interest expression feature.

[0096] In one example, taking the preset task goals including click, interaction and viewing time as an example, three multi-type behavioral interest expression features and the first material feature are input into the multi-head target attention mechanism model, and the output can be expressed as multipleinterest_click (i.e., the sub-multi-type behavioral interest expression feature for click behavior), multipleinterest_interact (i.e., the sub-multi-type behavioral interest expression feature for interactive behavior) and multipleinterest_duration (i.e., the sub-multi-type behavioral interest expression feature for viewing time).

[0097] In some embodiments, taking the preset task goals including clicks, interactions, and viewing time as an example, the multi-head target attention mechanism module splits the multi-type behavior interest expression feature into three sub-multi-type behavior interest expression features, namely the first multi-type behavior interest expression feature (corresponding to click behavior), the second sub-multi-type behavior interest expression feature (corresponding to interaction behavior), and the third sub-multi-type behavior interest expression feature (corresponding to viewing time). Then, the multi-head target attention mechanism module performs function mapping processing on the first material feature and the three sub-multi-type behavior interest expression features. Among them, in the process of the multi-head target attention mechanism module performing function mapping processing on the first material feature and the first sub-multi-type behavior interest expression feature, the parameter weight of the first network layer in the multi-head target attention mechanism module is expressed as W qc ∈R h×d ; The parameter weight of the second network layer in the multi-head target attention mechanism module is expressed as W kc ∈R h×d ; The parameter weight of the third network layer in the multi-head target attention mechanism module is expressed as W vc ∈R h×d ; The parameter weight of the fourth network layer in the multi-head target attention mechanism module is expressed as W mc ∈R h×1 ; Wherein, h and d are the dimensions of the network layer in the multi-head target attention mechanism module. In the process of the multi-head target attention mechanism module performing function mapping processing on the first material feature and the first sub-multi-type behavior interest expression feature, the function used is as follows:

[0098]

[0099] In formula (1), Q c represents the query vector of the multi-head target attention mechanism module, mid is the material ID of the candidate material, K c represents the key vector of the multi-head target attention mechanism module, mid_seq is the material ID sequence of the candidate material, V c A represents the numerical vector of the multi-head target attention mechanism module, multi_act_seq is the historical behavior information of the target user on the historical recommended materials, c is the preset weight, MutipleInterest c is the attention matrix function for click behavior, K c T K C The transposed matrix, d c The default dimensions.

[0100] In the process of the multi-head target attention mechanism module performing function mapping processing on the first material feature and the second sub-multi-type behavior interest expression feature, the parameter weight of the first network layer in the multi-head target attention mechanism module is expressed as W qi ∈R h×d ; The parameter weight of the second network layer in the multi-head target attention mechanism module is expressed as W ki ∈R h×d ; The parameter weight of the third network layer in the multi-head target attention mechanism module is expressed as W vi ∈R h×d ; The parameter weight of the fourth network layer in the multi-head target attention mechanism module is expressed as W mi ∈R h×1 ; Wherein, h and d are the dimensions of the network layer in the multi-head target attention mechanism module. In the process of the multi-head target attention mechanism module performing function mapping processing on the first material feature and the second sub-multi-type behavior interest expression feature, the function used is as follows:

[0101]

[0102] In formula (2), Q i represents the query vector of the multi-head target attention mechanism module, mid is the material ID of the candidate material, K i represents the key vector of the multi-head target attention mechanism module, mid_seq is the material ID sequence of the candidate material, V i A represents the numerical vector of the multi-head target attention mechanism module, multi_act_seq is the historical behavior information of the user on the historical recommended materials, i is the preset weight, MutipleInterest i is the attention matrix function for click behavior, K i T K i The transposed matrix, d i The default dimensions.

[0103] In the process of the multi-head target attention mechanism module performing function mapping processing on the first material feature and the third sub-multi-type behavior interest expression feature, the parameter weight of the first network layer in the multi-head target attention mechanism module is expressed as W qd ∈R h×d ; The parameter weight of the second network layer in the multi-head target attention mechanism module is expressed as W kd ∈R h×d ; The parameter weight of the third network layer in the multi-head target attention mechanism module is expressed as W vd ∈R h×d ; The parameter weight of the fourth network layer in the multi-head target attention mechanism module is expressed as W md ∈Rh×1 ; Wherein, h and d are the dimensions of the network layer in the multi-head target attention mechanism module. In the process of the multi-head target attention mechanism module performing function mapping processing on the first material feature and the third sub-multi-type behavior interest expression feature, the functions used are as follows:

[0104]

[0105] In formula (3), Q d represents the query vector of the multi-head target attention mechanism module, mid is the material ID of the candidate material, K d represents the key vector of the multi-head target attention mechanism module, mid_seq is the material ID sequence of the candidate material, V d A represents the numerical vector of the multi-head target attention mechanism module, multi_act_seq is the historical behavior information of the user on the historical recommended materials, d is the preset weight, MutipleInterest d is the attention matrix function for click behavior, K d T K d The transposed matrix, d d The default dimensions.

[0106] In some possible implementations, the above S302 may include: fusing the user feature and the second material feature through a sharing module to obtain a shared feature.

[0107] It can be understood that the shared module in the multi-type behavior interest model and the multi-head target attention mechanism module are in the same layer of network structure. The shared module in the multi-type behavior interest model can encapsulate commonly used functions or components, making it convenient for other modules in the multi-type behavior interest model to directly reference the data input to the shared module, reducing the time and cost of repeated development. The shared processing of the input data by the shared module is the reuse of the input data.

[0108] In some embodiments, the multi-layer perceptron network includes a plurality of sub-multi-layer perceptron networks, each of which corresponds to a task goal. After obtaining the multi-type behavioral interest expression features and shared features, the recommendation system can perform function mapping processing on the multi-type behavioral interest expression features and shared features through the multi-layer perceptron network to obtain the predicted value of the probability that the target user performs multiple types of user behaviors on the candidate material. Taking the preset task goals including: clicks, interactions, and viewing time as an example, the multi-layer perceptron network includes a first sub-multi-layer perceptron network corresponding to click behavior, a second sub-multi-layer perceptron network corresponding to interactive behavior, and a third sub-multi-layer perceptron network corresponding to viewing time.

[0109] In some embodiments, taking the preset task goals including clicks, interactions and viewing time as an example, the multi-layer perceptron network performs function mapping processing on the input multi-type behavioral interest expression features and sharing features, and the output prediction values ​​correspond to the user behaviors, namely: click_predic=MLP([share_feat,multipleinterest_click]), interact_predict=MLP([share_feat, multipleinterest_interact]) and duration_predic=MLP([share_feat,multipleinterest_duration]), wherein click_predic is the predicted value of the probability that the target user performs a click behavior, interact_predict is the predicted value of the probability that the target user performs an interactive behavior, and duration_predic is the predicted value of the viewing time of the target user for viewing the candidate material.

[0110] In some possible implementations, after executing S305, the material recommendation method may further include: collecting multiple types of user behaviors of the target user on the newly exposed target material to obtain real-time user behavior; encoding and integer converting the real-time user behavior to obtain the real-time behavior mark of the target user, and storing the real-time behavior mark in a graph storage database; generating training samples based on the material information of the target material, the user information of the target user and the real-time behavior mark, using the training samples to perform online training on a multi-type behavior interest model, adjusting the model parameters of the multi-type behavior interest model, and obtaining an updated multi-type behavior interest model.

[0111] It can be understood that after the recommendation system recommends the target material (i.e., the material selected from the candidate materials based on the prediction results) to the target user, the target user performs corresponding user behaviors on the target material according to his or her own interest level. The recommendation system collects the real-time user behaviors performed by the security user on the target material. By encoding and integer converting the real-time user behaviors, real-time behavior tags can be obtained. The real-time behavior tags, the material information of the target material, and the user information of the target user can be used together as training data to perform online training on the multi-type behavior interest model to obtain an updated multi-type behavior interest model.

[0112] In some embodiments, taking the preset task goals including clicks, interactions, and viewing time as an example, in the training process of the multi-type behavioral interest model, clicks and interactions are binary classification tasks, and the binary cross entropy loss function can be used; the viewing time is a regression task, and the mean square error can be used as the loss function. The final loss function is as follows:

[0113] Loss=BceLoss_click+BceLoss_interact+MseLoss_duration (4)

[0114] In formula (4), Loss is the loss function of the multi-type behavioral interest model, BceLoss_click is the parameter for click behavior, BceLoss_interact is the parameter for interactive behavior, and MseLoss_duration is the parameter for viewing time.

[0115] In some possible implementations, the above S305 may include: sorting the candidate materials based on the prediction results of the candidate materials under each task goal; and recommending the target material to the target user from the candidate materials based on the sorting results.

[0116] It can be understood that, taking the preset task objectives including clicks, interactions and viewing time as an example, the prediction results include three prediction values, corresponding to clicks, interactions and viewing time respectively. The recommendation system can sum the three prediction values ​​corresponding to each candidate material to obtain the total prediction value; then sort the total prediction values ​​of all candidate materials, and recommend the candidate materials whose sorting results meet the preset sorting threshold as target materials to the target users.

[0117] Figure 4 It is an optional structural diagram of the multi-type behavior interest model provided in the embodiment of the present application. Figure 5 This is an optional flow chart of the material recommendation method provided in the embodiment of the present application. Figure 4 right Figure 5 The steps shown are explained.

[0118] Below, the material recommendation method in the embodiment of the present application is explained by taking a multi-type behavioral interest model including a sharing module, a multi-head target attention mechanism module and a multi-layer perceptron network, and the preset task goals including: clicks, interactions and viewing time as an example.

[0119] S501, user A sends a recommendation request to the recommendation system.

[0120] S502, based on the recommendation request, the recommendation system obtains the historical behavior mark of user A from the user behavior graph database 51, and obtains the material information of the candidate material and the user information of user A from the user and material information database 52. The material information includes the material identification and the material content.

[0121] S503, the recommendation system processes the material identification of the candidate material, the historical behavior mark of user A, the user information of user A and the material content of the candidate material through the embedding layer to obtain the first material feature, the multi-type behavior sequence feature, the user feature and the second material feature.

[0122] S504, the recommendation system inputs the first material feature, the user feature, the second material feature and the multi-type behavior sequence feature into the multi-type behavior interest model 53 to obtain a prediction result.

[0123] Among them, in step S504, the sharing module 31 of the multi-type behavior interest model 53 performs a fusion process on the user features, the second material features and the first material features, and outputs the shared features. The multi-head target attention mechanism module 32 of the multi-type behavior interest model 53 performs a function mapping process on the multi-type behavior sequence features and the first material features to obtain the first sub-multi-type behavior sequence features (for click behavior), the second sub-multi-type behavior sequence features (for interactive behavior) and the third sub-multi-type behavior sequence features (for viewing time). The recommendation system combines the shared features with the first sub-multi-type behavior sequence features as a representation related to the click behavior, and inputs them into the first sub-multi-layer perceptron network 33-1 in the multi-layer perceptron network 33. The recommendation system combines the shared features with the second sub-multi-type behavior sequence features as a representation related to the interactive behavior, and inputs them into the second sub-multi-layer perceptron network 33-2 in the multi-layer perceptron network 33. The recommendation system combines the shared features with the third sub-multi-type behavior sequence features as a representation related to the viewing time, and inputs them into the third sub-multi-layer perceptron network 33-3 in the multi-layer perceptron network 33. The first sub-multi-layer perceptron network 33-1 performs function mapping processing on the input data to obtain a predicted value of the probability that user A will click on the candidate material; the second sub-multi-layer perceptron network 33-2 performs function mapping processing on the input data to obtain a predicted value of the probability that user A will interact with the candidate material; the third sub-multi-layer perceptron network 33-3 performs function mapping processing on the input data to obtain a predicted value of the probability that user A will watch the candidate material for a certain length of time.

[0124] S505, the recommendation system sorts the candidate materials according to the prediction results.

[0125] S506, the recommendation system selects the target material from the candidate materials according to the ranking results of the candidate materials, and performs user exposure (ie, recommends the target material to user A).

[0126] S507, the recommendation system collects the real-time user behavior of user A on the target material.

[0127] S508, the recommendation system pre-processes the real-time user behavior to obtain the real-time behavior tag, stores the real-time behavior tag in the user behavior graph database, and updates the user behavior graph database.

[0128] S509, the recommendation system generates training samples based on the real-time behavior mark of user A, the user characteristics of user A, and the material characteristics of the target material.

[0129] S510, the recommendation system trains a multi-type behavior interest model based on the training samples to obtain an updated multi-type behavior interest model 53.

[0130] At this point, the recommendation system has completed the above-mentioned material recommendation method.

[0131] In an embodiment of the present application, through the multi-head target attention mechanism module of the multi-type behavior interest model, based on the material characteristics and the multi-type behavior sequence characteristics, the multi-type behavior interest expression characteristics corresponding to each task goal are determined, and various types of user behaviors can be processed in a unified manner, thereby more accurately representing the target user's interest preference for the material.

[0132] Furthermore, based on various types of user behaviors, it is predicted that the target user will perform various types of user behaviors on the candidate materials, so that the prediction results are more in line with the user's interest preferences for the materials, and more accurate material recommendations are achieved.

[0133] Furthermore, by introducing encoding processing and integer conversion processing to obtain historical behavior tags, it is possible to encode multiple types of user behaviors with one tag, reduce the complexity of separate storage of multiple types of user behavior data, and reduce the complexity of data processing and modeling in the material recommendation process.

[0134] Furthermore, the encoded historical behavior markers are uniformly embedded into vectors (i.e., feature extraction processing), which can model the superposition and reinforcement relationship between different types of user behaviors and the behavioral dependency relationship for different task objectives.

[0135] Furthermore, based on the real-time user behavior of target users towards newly exposed target materials, online training of multi-type behavioral interest models can better express the real-time interest changes of target users.

[0136] The following is a description of an exemplary structure of the material recommendation device 455 provided in the embodiment of the present application implemented as a software module. In some embodiments, for example, Figure 2As shown, the software modules stored in the material recommendation device 455 of the memory 450 may include: a data acquisition module 4551, which is used to acquire the material characteristics of the candidate material, the user characteristics of the target user and the multi-type behavior sequence characteristics based on the recommendation request sent by the target user; the multi-type behavior sequence characteristics are used to characterize the target user's various types of user behaviors on multiple exposed materials within a set time range; a first data processing module 4552, which is used to obtain shared characteristics based on user characteristics and material characteristics through a sharing module of a multi-type behavior interest model; the shared characteristics are fusion characteristics of user characteristics and material characteristics; a second data processing module 4553, used to determine the multi-type behavior interest expression features corresponding to each task goal based on material features and multi-type behavior sequence features through the multi-head target attention mechanism module of the multi-type behavior interest model; the multi-type behavior interest expression features are used to characterize the target user's interest preference for materials under the corresponding task goals; the prediction module 4554, used to obtain the target user's prediction results for the candidate materials for the corresponding task goals based on each multi-type behavior interest expression feature and shared features through the multi-layer perceptron network of the multi-type behavior interest model; the recommendation module 4555, used to recommend materials to the target user based on the prediction results.

[0137] In some possible implementations, the above-mentioned device also includes: a feature extraction module, which is used to respond to a recommendation request sent by a target user, obtain material information of a candidate material, user information of a target user, and historical behavior marks of a target user within a set time range; perform feature extraction processing on the material information to obtain material features; perform feature extraction processing on the user information to obtain user features; perform feature extraction processing on the historical behavior marks to obtain multi-type behavior sequence features.

[0138] In some possible implementations, the above-mentioned device also includes: a behavior tag generation module, which is used to collect the historical behavior information of the target user on multiple exposed materials within a set time range in the target user's behavior log; encode the historical behavior information to obtain a binary array having a mapping relationship with the historical behavior information; perform integer conversion processing on the binary array to obtain a historical behavior tag having a mapping relationship with the binary array, and store the historical behavior tag in a graph storage database.

[0139] In some possible implementations, the material information includes first material information and second material information, the first material information is used to characterize the material code of the candidate material, and the second material information is used to characterize the material content of the candidate material; a feature extraction module is used to perform feature extraction processing on the first material information to obtain a first material feature; perform feature extraction processing on the second material information to obtain a second material feature; and combine the first material feature and the second material feature into a material feature.

[0140] In some possible implementations, the second data processing module 4553 is used to extract sub-multi-type behavior interest expression features corresponding to each task target from the multi-type behavior interest expression features through a multi-head target attention mechanism module; perform function mapping processing on each sub-multi-type behavior interest expression feature and the first material feature to obtain the multi-type behavior interest expression features corresponding to each task target.

[0141] In some possible implementations, the first data processing module 4552 is used to fuse the user features and the second material features through a sharing module to obtain shared features.

[0142] In some possible implementations, the recommendation module 4555 is used to sort the candidate materials based on the prediction results of the candidate materials under each task goal; and recommend the target material to the target user from the candidate materials based on the sorting results.

[0143] In some possible implementations, the above-mentioned device also includes: a training module, which is used to collect various types of user behaviors of the target user on the newly exposed target material to obtain real-time user behavior; encode the real-time user behavior and perform integer conversion processing to obtain the real-time behavior mark of the target user, and store the real-time behavior mark in a graph storage database; generate training samples based on the material information of the target material, the user information of the target user and the real-time behavior mark, use the training samples to perform online training on the multi-type behavior interest model, adjust the model parameters of the multi-type behavior interest model, and obtain an updated multi-type behavior interest model.

[0144] The embodiment of the present application provides a computer program product or a computer program, which includes computer instructions, and the computer instructions are stored in a computer storage medium. The processor of the computer device reads the computer instructions from the computer storage medium, and the processor executes the computer instructions, so that the computer device executes the material recommendation method mentioned in the embodiment of the present application.

[0145] The present application embodiment provides a computer storage medium storing executable instructions, wherein the executable instructions are stored. When the executable instructions are executed by a processor, the processor will execute the material recommendation method provided by the present application embodiment, for example Figure 3 Recommended method for materials shown.

[0146] In some embodiments, the computer storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or it may be various devices including one or any combination of the above memories.

[0147] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), 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.

[0148] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores 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 storing one or more modules, subroutines, or code portions).

[0149] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0150] The above are only embodiments of the present application and are not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A material recommendation method, characterized in that: The method comprises: Based on the recommendation request sent by the target user, the material characteristics of the candidate material, the user characteristics of the target user and the multi-type behavior sequence characteristics are obtained; the multi-type behavior sequence characteristics are used to characterize the multiple types of user behaviors of the target user on multiple exposed materials within a set time range; A shared feature is obtained based on the user feature and the material feature through a shared module of a multi-type behavior interest model; the shared feature is a fusion feature of the user feature and the material feature; Through the multi-head target attention mechanism module of the multi-type behavior interest model, based on the material features and the multi-type behavior sequence features, the multi-type behavior interest expression features corresponding to each task target are determined; the multi-type behavior interest expression features are used to characterize the interest preference of the target user for the material under the corresponding task target; Obtaining the prediction result of the target user on the candidate material for the corresponding task goal based on each multi-type behavior interest expression feature and the shared feature through the multi-layer perceptron network of the multi-type behavior interest model; Based on the prediction result, material recommendations are made to the target user.

2. The method according to claim 1, characterized in that: Before acquiring the material features of the candidate materials, the user features of the target users, and the multi-type behavior sequence features, the method further includes: In response to a recommendation request sent by the target user, obtaining material information of the candidate material, user information of the target user, and a historical behavior mark of the target user within the set time range; Performing feature extraction processing on the material information to obtain the material features; Performing feature extraction processing on the user information to obtain the user features; The historical behavior marks are subjected to feature extraction processing to obtain the multi-type behavior sequence features.

3. The method according to claim 2, characterized in that Before obtaining the historical behavior mark of the target user within the set time range, the method further includes: Collecting historical behavior information of the target user on the multiple exposed materials within the set time range from the behavior log of the target user; Encoding the historical behavior information to obtain a binary array having a mapping relationship with the historical behavior information; An integer conversion process is performed on the binary array to obtain a historical behavior mark having a mapping relationship with the binary array, and the historical behavior mark is stored in a graph storage database.

4. The method according to claim 2, characterized in that: The material information includes first material information and second material information, the first material information is used to represent the material code of the candidate material, and the second material information is used to represent the material content of the candidate material; The performing feature extraction processing on the material information to obtain the material features includes: Performing feature extraction processing on the first material information to obtain first material features; Performing feature extraction processing on the second material information to obtain second material features; The first material characteristic and the second material characteristic are combined into the material characteristic.

5. The method according to claim 4, characterized in that The multi-head target attention mechanism module of the multi-type behavior interest model determines the multi-type behavior interest expression features corresponding to each task target based on the material features and the multi-type behavior sequence features, including: Extracting sub-multi-type behavior interest expression features corresponding to each task goal from the multi-type behavior interest expression features through the multi-head target attention mechanism module; Function mapping processing is performed on each sub-multi-type behavior interest expression feature and the first material feature to obtain the multi-type behavior interest expression feature corresponding to each task target.

6. The method according to claim 4, characterized in that The shared features are obtained based on the user features and the material features through the shared module of the multi-type behavior interest model, including: The user feature and the second material feature are fused through the sharing module to obtain the shared feature.

7. The method according to claim 1, characterized in that The recommending materials to the target user based on the prediction result includes: Based on the prediction results of the candidate materials under each task objective, the candidate materials are sorted; Based on the ranking result, a target material is recommended to the target user from the candidate materials.

8. The method according to claim 1, characterized in that After recommending materials to the target user based on the prediction result, the method further includes: Collecting various types of user behaviors of the target user on the newly exposed target material to obtain real-time user behaviors; Performing encoding and integer conversion processing on the real-time user behavior to obtain a real-time behavior tag of the target user, and storing the real-time behavior tag in a graph storage database; A training sample is generated based on the material information of the target material, the user information of the target user and the real-time behavior mark, the multi-type behavior interest model is trained online using the training sample, the model parameters of the multi-type behavior interest model are adjusted, and an updated multi-type behavior interest model is obtained.

9. A material recommendation device, characterized in that: The device comprises: A data acquisition module, used to acquire material features of candidate materials, user features of the target user, and multi-type behavior sequence features based on a recommendation request sent by a target user; the multi-type behavior sequence features are used to characterize the target user's multi-type user behaviors on multiple exposed materials within a set time range; A first data processing module is used to obtain a shared feature based on the user feature and the material feature through a sharing module of a multi-type behavior interest model; the shared feature is a fusion feature of the user feature and the material feature; A second data processing module is used to determine the multi-type behavior interest expression features corresponding to each task target based on the material features and the multi-type behavior sequence features through the multi-head target attention mechanism module of the multi-type behavior interest model; the multi-type behavior interest expression features are used to characterize the interest preference of the target user for the material under the corresponding task target; A prediction module, configured to obtain a prediction result of the target user on the candidate material for a corresponding task goal based on each multi-type behavior interest expression feature and the shared feature through a multi-layer perceptron network of the multi-type behavior interest model; A recommendation module is used to recommend materials to the target user based on the prediction results.

10. An electronic device, characterized in that: The electronic device comprises: A memory for storing executable instructions; A processor, configured to implement the method according to any one of claims 1 to 8 when executing the executable instructions or computer programs stored in the memory.

11. A computer-readable storage medium storing executable instructions or a computer program, characterized in that: When the executable instructions are executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Citation Information

Cited By

  • Information recommendation method, model training method, equipment, storage medium and program product

    CN120163622A