Community material recommendation method, device and equipment based on large model and storage medium

Abstract

The invention provides a community material recommendation method and device based on a large model, equipment and a storage medium, and the method comprises the steps: obtaining the personal information of a user, converting the personal information features of the user into user feature vectors through an embedded layer, obtaining the historical behavior data of the user, processing the historical behavior data through a pre-trained large model, and carrying out the processing of the historical behavior data. A user behavior feature vector is generated, a weight is allocated to each material through an activation function, feature information of candidate materials is determined and obtained according to the weight of each material, and the feature information of the candidate materials is converted into candidate material feature vectors through an embedded layer; and splicing the user feature vector, the user behavior feature vector and the candidate material feature vector to generate a spliced vector, calculating the spliced vector through a full connection layer and a nonlinear activation function to generate an interest weight of the user for the candidate material, and outputting a recommendation result generated based on the interest weight through Softmax layer sorting. The problem that effective material pushing cannot be carried out on users based on multi-modal data in a community is solved.

Description

Technical Field

[0001] The present invention relates to the field of computers, and in particular to a community material recommendation method, device, equipment and storage medium based on a large model. Background Art

[0002] In today's era of information explosion, users are faced with a vast amount of content to choose from, which poses a huge challenge to personalized recommendation systems. Traditional recommendation algorithms mainly rely on users' historical behavior data and simple feature engineering, but this approach is incapable of processing complex, multimodal data. In particular, when processing unstructured data such as user comments, conversations, pictures, and videos, traditional methods find it difficult to fully extract and utilize the rich semantic information contained therein.

[0003] In view of this, this application is filed. Summary of the invention

[0004] The present invention discloses a community material recommendation method, device, equipment and storage medium based on a large model, aiming to solve the problem that it is impossible to effectively push materials to users based on multimodal data in the community.

[0005] The first embodiment of the present invention provides a community material recommendation method based on a large model, comprising:

[0006] Obtain the user's personal information and convert the user's personal information features into a user feature vector through the embedding layer;

[0007] Obtain the user's historical behavior data, and process the historical behavior data through a pre-trained large model to generate a user behavior feature vector, where the user's historical behavior data includes the user's interaction record with each material;

[0008] Assign a weight to each material through an activation function, determine the feature information of the candidate material according to the weight of each material, and convert the feature information of the candidate material into a candidate material feature vector through an embedding layer, wherein the weight can be dynamically adjusted based on user behavior;

[0009] The user feature vector, user behavior feature vector, and candidate material feature vector are concatenated to generate a concatenated vector, the concatenated vector is calculated through a fully connected layer and a nonlinear activation function to generate the user's interest weight for the candidate material, and the recommendation result generated based on the interest weight is sorted and output through a Softmax layer.

[0010] Preferably, the obtaining of the user's personal information and converting the user's personal information into a user feature vector through an embedding layer is specifically as follows:

[0011] Receive the user's personal information, generate the user's personality tag through prompt engineering and large language model, and convert the personality tag into a user feature vector, wherein the personal information includes registration information, conversation data and comment data.

[0012] Preferably, the user's historical behavior data is obtained and processed by a pre-trained large model to generate a user behavior feature vector, specifically:

[0013] Extract the user's interaction records with the material in the target community, and call the pre-trained large model to jointly process and extract features of the material information to generate feature vectors of different modalities, wherein the interaction records with the material include video information, text information, and audio information;

[0014] The feature vectors of the different modalities are mapped to the same semantic space, and the features of the different modalities are adaptively weighted through an attention mechanism or a gating mechanism, and the weighted features are fused to generate a user behavior feature vector.

[0015] Preferably, it also includes:

[0016] Obtain the user's registration information and generate a temporary tag for the user based on the first prompt project;

[0017] The temporary labels and candidate material lists are converted into sorting tasks through the second prompting project, and the sorting results of the candidate materials are generated based on the large language model.

[0018] The second embodiment of the present invention provides a community material recommendation device based on a large model, comprising:

[0019] A personal information processing unit, used to obtain the user's personal information and convert the user's personal information features into a user feature vector through an embedding layer;

[0020] A behavior data processing unit is used to obtain the user's historical behavior data and process the historical behavior data through a pre-trained large model to generate a user behavior feature vector, wherein the user's historical behavior data includes the user's interaction record with each material;

[0021] A weight assignment unit, used to assign a weight to each material through an activation function, determine and obtain feature information of the candidate material according to the weight of each material, and convert the feature information of the candidate material into a feature vector of the candidate material through an embedding layer, wherein the weight can be dynamically adjusted based on user behavior;

[0022] The recommendation result generating unit is used to concatenate the user feature vector, the user behavior feature vector, and the candidate material feature vector to generate a concatenated vector, calculate the concatenated vector through a fully connected layer and a nonlinear activation function to generate the user's interest weight for the candidate material, and sort and output the recommendation result generated based on the interest weight through a Softmax layer.

[0023] Preferably, the personal information processing unit is specifically used for:

[0024] Receive the user's personal information, generate the user's personality tag through prompt engineering and large language model, and convert the personality tag into a user feature vector, wherein the personal information includes registration information, conversation data and comment data.

[0025] Preferably, the behavior data processing unit is used for:

[0026] Extract the user's interaction records with the material in the target community, and call the pre-trained large model to jointly process and extract features of the material information to generate feature vectors of different modalities, wherein the interaction records with the material include video information, text information, and audio information;

[0027] The feature vectors of the different modalities are mapped to the same semantic space, and the features of the different modalities are adaptively weighted through an attention mechanism or a gating mechanism, and the weighted features are fused to generate a user behavior feature vector.

[0028] Preferably, the personal information processing unit also communicates with the user:

[0029] Obtain the user's registration information and generate a temporary tag for the user based on the first prompt project;

[0030] The temporary labels and candidate material lists are converted into sorting tasks through the second prompting project, and the sorting results of the candidate materials are generated based on the large language model.

[0031] The third embodiment of the present invention provides a community material recommendation device based on a large model, including a memory and a processor, wherein a computer program is stored in the memory, and the computer program can be executed by the processor to implement a community material recommendation method based on a large model as described in any one of the above.

[0032] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program can be executed by a processor of a device where the computer-readable storage medium is located to implement a community material recommendation method based on a large model as described in any one of the above items.

[0033] A method, device, equipment and storage medium for community material recommendation based on a large model provided by the present invention first obtains the user's personal information, and converts the user's personal information features into a user feature vector through an embedding layer, then obtains the user's historical behavior data, and processes the historical behavior data through a pre-trained large model to generate a user behavior feature vector, wherein the user's historical behavior data includes the user's interaction record with each material; then, a weight is assigned to each material through an activation function, and the feature information of the candidate material is obtained according to the weight of each material, and the feature information of the candidate material is converted into a candidate material feature vector through an embedding layer, wherein the weight can be dynamically adjusted based on the user's behavior; finally, the user feature vector, the user behavior feature vector, and the candidate material feature vector are spliced ​​to generate a splicing vector, and the splicing vector is calculated through a fully connected layer and a nonlinear activation function to generate the user's interest weight for the candidate material, and the recommendation result generated based on the interest weight is output through a Softmax layer sorting. The problem that it is impossible to effectively push materials to users based on multimodal data in the community is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flowchart of a community material recommendation method based on a large model provided by the first embodiment of the present invention;

[0035] Figure 2 It is a module diagram of a community material recommendation device based on a large model provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] In order to better understand the technical solution of the present invention, the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0038] It should be clear that the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0040] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0041] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0042] The "first\second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order for the objects. It is understandable that the "first\second" can be interchanged with the specific order or sequence where permitted. It should be understood that the objects distinguished by "first\second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0043] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0044] The present invention discloses a community material recommendation method, device, equipment and storage medium based on a large model, aiming to solve the problem that it is impossible to effectively push materials to users based on multimodal data in the community.

[0045] See also Figure 1 The first embodiment of the present invention provides a community material recommendation method based on a large model, which can be executed by a recommendation device based on a large model (hereinafter referred to as a recommendation device), and in particular, executed by one or more processors in the recommendation device to at least implement the following steps:

[0046] S101, obtaining the user's personal information, and converting the user's personal information features into a user feature vector through an embedding layer;

[0047] In this embodiment, the recommended device may be a terminal with data processing capability, such as a server, a workstation, a desktop computer, a laptop computer, a tablet computer, etc., and the recommended device may be installed with a corresponding operating system and application software, and the functions required by this embodiment may be realized through the combination of the operating system and the application software;

[0048] Specifically, in this embodiment, the user's personal information can be received first. The personal information may include but is not limited to the user's registration information, the user's conversation data, and the user's comment data. The registration information covers the basic information registered by the user in the system, such as name, age, gender, occupation, etc. The conversation data is the text or voice conversation content generated during the interaction between the user and the system, and the user comment data refers to the evaluation and feedback of other materials or content published by the user on the community.

[0049] The recommendation device can process this information using prompt engineering. The purpose of prompt engineering is to provide accurate input prompts for the large language model (LLM) to ensure that the LLM can understand and accurately process the acquired user information. In prompt engineering, the recommendation device sets corresponding prompt templates for each type of personal information (such as registration information, conversation data, comment data, etc.). For example, the recommendation device can generate a structured prompt for the registration information to inform the LLM that this is the user's basic information; while for the conversation data and comment data, it will generate content analysis prompts to help the LLM identify potential interests and preferences in this information. Furthermore, the LLM will combine the information in the user's registration information, conversation data, and comment data, and use natural language processing technology to mine the user's behavior patterns, areas of interest, and personality characteristics. For example, by analyzing the keywords in the conversation data and comment data, the LLM can infer the user's emotional tendencies and preferences for a certain type of material in a specific time period.

[0050] After analyzing the user information, the large language model will generate a personalized label. The personalized label can reflect the user's personalized characteristics, such as the user's long-term interest in a certain type of goods or services, consumption habits, favorite content types, etc. These personalized labels are not simple static descriptions, but dynamic labels formed by combining the user's multi-dimensional information, which can be updated over time and with behavioral changes.

[0051] The generated personalized tags are further converted into user feature vectors. Specifically, the recommendation device uses a preset embedding model to convert the personalized tags into corresponding numerical representations, which are user feature vectors. User feature vectors are low-dimensional vector representations used to characterize the personalized information of users in specific scenarios.

[0052] S102, obtaining historical behavior data of the user, and processing the historical behavior data through a pre-trained large model to generate a user behavior feature vector, wherein the historical behavior data of the user includes a record of interaction between the user and each material;

[0053] In this embodiment, the interaction records between users and materials in the target community can be extracted. The interaction records include but are not limited to video information, text information and audio information related to users and materials. These interaction records can reflect the interaction between users and different contents in the community, such as videos watched by users, articles read, text content when providing comments or feedback, or behavioral data generated during the interaction with audio content. Further, the recommendation device calls a pre-trained large model to process the extracted material information. The large model is trained in multi-modal joint and can extract features from information in different forms. Specifically, video information, text information and audio information will be encoded by corresponding processing modules respectively. Video information can be extracted by image features through a pre-trained convolutional neural network (CNN), text information can be understood and encoded by a large model based on Transformer, and audio information is processed by an acoustic model. After processing, the recommendation device can generate feature representations for data of each modality, that is, corresponding video feature vectors, text feature vectors and audio feature vectors.

[0054] Furthermore, the recommendation device maps the feature vectors of these different modalities to the same semantic space. The purpose of this step is to ensure that data from different modalities can be compared and operated under the same representation dimension. That is, the video, text and audio features are semantically aligned. In order to further improve the accuracy and adaptability of feature fusion, the recommendation device introduces an attention mechanism or a gating mechanism. It can dynamically adjust the weight of each modal feature according to the importance of each modality. For example, the attention mechanism can give a higher weight to the features of a modality based on the user's attention to the modality, and give a lower weight to the modality with less attention. The weighted feature vectors are further fused. The recommendation device generates a unified user behavior feature vector by fusing the weighted features of different modalities. The user behavior feature vector not only integrates the interaction records between the user and the material, but also reflects the user's comprehensive interests and behavior patterns under multimodal information.

[0055] S103, assigning a weight to each material through an activation function, determining feature information of the candidate material according to the weight of each material, and converting the feature information of the candidate material into a feature vector of the candidate material through an embedding layer, wherein the weight can be dynamically adjusted based on user behavior;

[0056] It should be noted that in this embodiment, the recommendation device receives historical behavior data from the user and analyzes the interaction records between the user and each material. The interaction records include user clicks, browses, purchases, and other behaviors. Based on these behaviors, the recommendation device can preliminarily determine the user's interest in different materials. Then, the recommendation device uses an activation function to assign an initial weight to each material, and the size of the weight depends on the interaction intensity between the user and the material. For example, for materials that users frequently click or purchase, the system will assign a higher weight, while for materials that users pay less attention to, the system will assign a lower weight.

[0057] The activation function can adopt common nonlinear functions, such as ReLU, Sigma, or other forms, to adjust the weight distribution of materials. In addition, the weight is not only static, but can be adjusted dynamically. Specifically, the recommendation device will provide real-time feedback and weight update according to the real-time behavior of the user. For example, when a user frequently browses certain materials in a short period of time, the system will increase the weight of these materials accordingly, and vice versa.

[0058] After assigning weights, the recommendation device determines the candidate material feature information that needs to be further processed based on the weight of each material. The recommendation device extracts the feature information of these candidate materials from the material database, which may include the material category, brand, price, user evaluation, etc. The extracted candidate material feature information is then processed by the embedding layer. The embedding layer is responsible for converting these high-dimensional material feature information into low-dimensional feature vectors, so that this information can be efficiently calculated and represented in the neural network.

[0059] S104, concatenating the user feature vector, the user behavior feature vector, and the candidate material feature vector to generate a concatenated vector, calculating the concatenated vector through a fully connected layer and a nonlinear activation function to generate the user's interest weight for the candidate material, and sorting and outputting the recommendation result generated based on the interest weight through a Softmax layer.

[0060] It should be noted that in this embodiment, the recommendation device concatenates the user feature vector, the user behavior feature vector, and the candidate material feature vector to generate a comprehensive concatenation vector. The user feature vector contains the user's basic information, such as age, gender, interest preferences, etc., while the user behavior feature vector is derived from the user's historical behavior data, reflecting the user's past interaction records and behavior patterns. The candidate material feature vector is obtained by embedding the material feature information, covering the material's attributes, categories, and other related information.

[0061] By concatenating these three types of feature vectors, the recommendation device combines the multi-dimensional information of users, user behaviors, and candidate materials in a unified representation space. The concatenated vector will contain all relevant information. At this time, the recommendation device inputs the concatenated vector into the fully connected layer, which performs weighted calculations on it. The fully connected layer linearly combines each dimensional feature in the concatenated vector through the learned weight matrix, thereby extracting the user's potential interest pattern in the candidate materials.

[0062] Next, the recommendation device introduces nonlinear activation functions, such as ReLU, Sigma, or PReLU, to perform nonlinear transformations on the results calculated by the fully connected layer. This captures more complex nonlinear relationships between users and materials, allowing the model to more accurately predict the user's interest in a specific candidate material. The output of the activation function represents the user's interest weight for each candidate material.

[0063] After generating the user's interest weights for the candidate materials, the recommendation device normalizes the interest weights of all candidate materials through the Softmax layer. The Softmax layer converts the interest weights of all candidate materials into a probability distribution, ensuring that the sum of the weights of each candidate material is 1. Intuitively, the weight of each candidate material is mapped to a recommendation probability, indicating which candidate materials the user is more likely to be interested in. Finally, the recommendation device sorts the candidate materials according to the output of the Softmax layer, giving priority to those with higher interest weights. The recommendation result not only fully considers the user's characteristics and behavior history, but also combines the characteristic information of the candidate materials to generate personalized and highly relevant recommendation content.

[0064] In a possible embodiment of the present invention, the method may further include:

[0065] Obtain the user's registration information and generate a temporary tag for the user based on the first prompt project;

[0066] The temporary labels and candidate material lists are converted into sorting tasks through the second prompting project, and the sorting results of the candidate materials are generated based on the large language model.

[0067] It should be noted that during the cold start phase, the recommendation device extracts key information from the user's registration information. These registration information usually include basic information about the user, such as age, gender, occupation, hobbies, etc., which can provide a reference for the user's initial profile. The recommendation device generates temporary tags for the user based on these registration information through a predefined first prompt project. The prompt project is a guidance process based on predefined rules or models that can combine the user's registration information with the tag system to generate initial personalized tags for each new user. These temporary tags reflect the user's basic attributes and potential interests. Even if the user does not provide a large amount of interaction data, the recommendation device can still generate preliminary recommendations for them through these tags.

[0068] After generating the user's temporary tags, the system will screen and process the candidate material list based on these tags. Through the second prompt project, the system matches the user's temporary tags with the characteristics of the candidate materials and generates a sorting task. This sorting task means that the recommendation device needs to sort the candidate materials according to the degree of match with the user's potential interests based on the user's initial tag information. In this process, the characteristic information of the candidate materials, such as category, audience, usage scenario, etc., will be taken into account to ensure that the recommended materials are consistent with the user's temporary tags.

[0069] To further improve the accuracy of recommendation results, the recommendation device can call a large language model to perform this sorting task. It can capture complex semantic associations in the process of analyzing the user's temporary tags and the characteristics of candidate materials. By understanding and reasoning about these tags and material characteristics, the large language model can generate personalized candidate material sorting results for each user, that is, putting the most relevant and most likely to arouse the user's interest at the top of the recommendation list.

[0070] See also Figure 2 The second embodiment of the present invention provides a community material recommendation device based on a large model, comprising:

[0071] The personal information processing unit 201 is used to obtain the user's personal information and convert the user's personal information features into a user feature vector through an embedding layer;

[0072] The behavior data processing unit 202 is used to obtain the user's historical behavior data and process the historical behavior data through the pre-trained large model to generate a user behavior feature vector, wherein the user's historical behavior data includes the interaction record between the user and each material;

[0073] A weight assignment unit 203 is used to assign a weight to each material through an activation function, determine and obtain feature information of the candidate material according to the weight of each material, and convert the feature information of the candidate material into a feature vector of the candidate material through an embedding layer, wherein the weight can be dynamically adjusted based on user behavior;

[0074] The recommendation result generating unit 204 is used to concatenate the user feature vector, the user behavior feature vector, and the candidate material feature vector to generate a concatenated vector, calculate the concatenated vector through a fully connected layer and a nonlinear activation function to generate the user's interest weight for the candidate material, and sort and output the recommendation result generated based on the interest weight through a Softmax layer.

[0075] Preferably, the personal information processing unit is specifically used for:

[0076] Receive the user's personal information, generate the user's personality tag through prompt engineering and large language model, and convert the personality tag into a user feature vector, wherein the personal information includes registration information, conversation data and comment data.

[0077] Preferably, the behavior data processing unit is used for:

[0078] Extract the user's interaction records with the material in the target community, and call the pre-trained large model to jointly process and extract features of the material information to generate feature vectors of different modalities, wherein the interaction records with the material include video information, text information, and audio information;

[0079] The feature vectors of the different modalities are mapped to the same semantic space, and the features of the different modalities are adaptively weighted through an attention mechanism or a gating mechanism, and the weighted features are fused to generate a user behavior feature vector.

[0080] Preferably, the personal information processing unit also communicates with the user:

[0081] Obtain the user's registration information and generate a temporary tag for the user based on the first prompt project;

[0082] The temporary labels and candidate material lists are converted into sorting tasks through the second prompting project, and the sorting results of the candidate materials are generated based on the large language model.

[0083] The third embodiment of the present invention provides a community material recommendation device based on a large model, including a memory and a processor, wherein a computer program is stored in the memory, and the computer program can be executed by the processor to implement a community material recommendation method based on a large model as described in any one of the above.

[0084] The fourth embodiment of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program can be executed by a processor of a device where the computer-readable storage medium is located to implement a community material recommendation method based on a large model as described in any one of the above items.

[0085] A method, device, equipment and storage medium for community material recommendation based on a large model provided by the present invention first obtains the user's personal information, and converts the user's personal information features into a user feature vector through an embedding layer, then obtains the user's historical behavior data, and processes the historical behavior data through a pre-trained large model to generate a user behavior feature vector, wherein the user's historical behavior data includes the user's interaction record with each material; then, a weight is assigned to each material through an activation function, and the feature information of the candidate material is obtained according to the weight of each material, and the feature information of the candidate material is converted into a candidate material feature vector through an embedding layer, wherein the weight can be dynamically adjusted based on the user's behavior; finally, the user feature vector, the user behavior feature vector, and the candidate material feature vector are spliced ​​to generate a splicing vector, and the splicing vector is calculated through a fully connected layer and a nonlinear activation function to generate the user's interest weight for the candidate material, and the recommendation result generated based on the interest weight is output through a Softmax layer sorting. The problem that it is impossible to effectively push materials to users based on multimodal data in the community is solved.

[0086] Exemplarily, the computer program described in the third and fourth embodiments of the present invention may be divided into one or more modules, and the one or more modules are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program in the implementation of a community material recommendation device based on a large model. For example, the device described in the second embodiment of the present invention.

[0087] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the community material recommendation method based on a large model, and uses various interfaces and lines to connect the various parts of the entire implementation of the community material recommendation method based on a large model.

[0088] The memory can be used to store the computer program and / or module, and the processor realizes various functions of a community material recommendation method based on a large model by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, a text conversion function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, text message data, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Med ia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0089] Wherein, if the implemented module is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0090] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.

[0091] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A community material recommendation method based on a large model, characterized in that: include: Obtain the user's personal information and convert the user's personal information features into a user feature vector through the embedding layer; Obtain the user's historical behavior data, and process the historical behavior data through a pre-trained large model to generate a user behavior feature vector, where the user's historical behavior data includes the user's interaction record with each material; Assign a weight to each material through an activation function, determine the feature information of the candidate material according to the weight of each material, and convert the feature information of the candidate material into a candidate material feature vector through an embedding layer, wherein the weight can be dynamically adjusted based on user behavior; The user feature vector, user behavior feature vector, and candidate material feature vector are concatenated to generate a concatenated vector, the concatenated vector is calculated through a fully connected layer and a nonlinear activation function to generate the user's interest weight for the candidate material, and the recommendation result generated based on the interest weight is sorted and output through a Softmax layer.

2. A community material recommendation method based on a large model according to claim 1, characterized in that: The method of obtaining the user's personal information and converting the user's personal information into a user feature vector through an embedding layer is specifically as follows: Receive the user's personal information, generate the user's personality tag through prompt engineering and large language model, and convert the personality tag into a user feature vector, wherein the personal information includes registration information, conversation data and comment data.

3. The method for community material recommendation based on a large model according to claim 1, characterized in that: The user's historical behavior data is obtained, and the historical behavior data is processed by a pre-trained large model to generate a user behavior feature vector, specifically: Extract the user's interaction records with the material in the target community, and call the pre-trained large model to jointly process and extract features of the material information to generate feature vectors of different modalities, wherein the interaction records with the material include video information, text information, and audio information; The feature vectors of the different modalities are mapped to the same semantic space, and the features of the different modalities are adaptively weighted through an attention mechanism or a gating mechanism, and the weighted features are fused to generate a user behavior feature vector.

4. A community material recommendation method based on a large model according to claim 2, characterized in that: Also includes: Obtain the user's registration information and generate a temporary tag for the user based on the first prompt project; The temporary labels and candidate material lists are converted into sorting tasks through the second prompting project, and the sorting results of the candidate materials are generated based on the large language model.

5. A community material recommendation device based on a large model, characterized in that: include: A personal information processing unit, used to obtain the user's personal information and convert the user's personal information features into a user feature vector through an embedding layer; A behavior data processing unit is used to obtain the user's historical behavior data and process the historical behavior data through a pre-trained large model to generate a user behavior feature vector, wherein the user's historical behavior data includes the user's interaction record with each material; A weight assignment unit, used to assign a weight to each material through an activation function, determine and obtain feature information of the candidate material according to the weight of each material, and convert the feature information of the candidate material into a feature vector of the candidate material through an embedding layer, wherein the weight can be dynamically adjusted based on user behavior; The recommendation result generating unit is used to concatenate the user feature vector, the user behavior feature vector, and the candidate material feature vector to generate a concatenated vector, calculate the concatenated vector through a fully connected layer and a nonlinear activation function to generate the user's interest weight for the candidate material, and sort and output the recommendation result generated based on the interest weight through a Softmax layer.

6. A community material recommendation device based on a large model according to claim 5, characterized in that: The personal information processing unit is specifically used for: Receive the user's personal information, generate the user's personality tag through prompt engineering and large language model, and convert the personality tag into a user feature vector, wherein the personal information includes registration information, conversation data and comment data.

7. The large model-based community material recommendation device according to claim 5, characterized in that: The behavior data processing unit is used for: Extract the user's interaction records with the material in the target community, and call the pre-trained large model to jointly process and extract features of the material information to generate feature vectors of different modalities, wherein the interaction records with the material include video information, text information, and audio information; The feature vectors of the different modalities are mapped to the same semantic space, and the features of the different modalities are adaptively weighted through an attention mechanism or a gating mechanism, and the weighted features are fused to generate a user behavior feature vector.

8. The large model-based community material recommendation device according to claim 6, characterized in that: The personal information processing unit also communicates with the user: Obtain the user's registration information and generate a temporary tag for the user based on the first prompt project; The temporary labels and candidate material lists are converted into sorting tasks through the second prompting project, and the sorting results of the candidate materials are generated based on the large language model.

9. A community material recommendation device based on a large model, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement a community material recommendation method based on a large model as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that: A computer program is stored, and the computer program can be executed by a processor of the device where the computer-readable storage medium is located to implement a community material recommendation method based on a large model as described in any one of claims 1 to 4.

Images