Multi-agent-based content recommendation method and device, equipment, medium and product

Abstract

The invention discloses a multi-agent-based content recommendation method and device, equipment, a medium and a product, and the method comprises the steps: inputting obtained user data into each pre-constructed recall agent, and obtaining recommendation content outputted by each recall agent; inputting the user data into a pre-constructed large language model to obtain a selection strategy of each recall agent; and according to the selection strategy, carrying out fusion processing on the recommendation content output by each recall agent to obtain fused recommendation content. According to the scheme, through fusion of the recommended contents of the multiple recall agents, a more comprehensive recommended content output scheme is provided, the scene application range is widened, content fusion is carried out based on the calculated selection strategy, and the content recommendation accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of information recommendation, and in particular to a multi-agent-based content recommendation method, device, equipment, medium and product. Background Art

[0002] With the development of information technology and Internet technology, people have gradually moved from an era of information scarcity to an era of information overload. In this case, consumers want to find information of interest from a large amount of information (such as items in online shopping malls), while information producers want their information to stand out. Information recommendation systems can connect users and information, predict users' ratings or preferences for items, and build a bridge between the two. Recommendation systems discover users' behaviors, find users' personalized needs, and help users find information that interests them.

[0003] Existing recommendation systems build user preference models by extracting information features and user past behaviors. They identify and recommend information that users may be interested in by comparing user preference models and item features. Existing multi-agent-based content recommendation methods are difficult to apply to cold start scenarios with insufficient data richness, and the accuracy of content recommendations is low. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a multi-agent based content recommendation method, device, equipment, medium and product to improve the scope of application of scenarios and the accuracy of content recommendation.

[0005] An embodiment of the present invention provides a multi-agent based content recommendation method, the method comprising:

[0006] Input the acquired user data into each pre-built recall agent respectively, and obtain the recommended content output by each recall agent;

[0007] Inputting the user data into a pre-built large language model to obtain a selection strategy for each recall agent;

[0008] The recommendation contents output by each recall agent are fused according to the selection strategy to obtain fused recommendation contents.

[0009] Preferably, the method further comprises:

[0010] Inputting the user data and the fused recommended content into a preset interpretation agent to obtain recommended interpretation content;

[0011] The recommended explanation content and the fused recommended content are output as a recommendation result.

[0012] As a preferred solution, the recommendation contents output by each recall agent are fused according to the selection strategy to obtain fused recommendation contents, including:

[0013] Take the union of the recommendation results output by each recall agent to get the overall recommendation content;

[0014] Calculating the comprehensive score of each material in the overall recommended content according to the selection strategy;

[0015] The fused recommended content is selected according to the comprehensive scores of the various materials in the overall recommended content and preset screening rules.

[0016] Furthermore, the score of each material in the overall recommended content is calculated according to the selection strategy and a preset scoring calculation model, including:

[0017] Taking the confidence in the selection strategy as a weight, weighted summing is performed on the recommendation scores of each material in the overall recommendation content in each recall agent to determine the recommendation score of each material in the overall recommendation content;

[0018] The recommendation scores of each item are input into a preset scoring calculation model, and the scores of each item in the overall recommended content in different dimensions are weighted to determine the comprehensive score of each item in the overall recommended content.

[0019] Preferably, the acquired user data is respectively input into each pre-built recall agent to obtain the recommended content output by each recall agent, including:

[0020] Input the acquired user data into each recall agent respectively to obtain the recommendation output of each recall agent;

[0021] Use a pre-trained vector model to convert each recommendation output into a vector representation in a high-dimensional space;

[0022] Each vector representation is used as a query statement, and compared with the material vectors in the preset vector database to determine the corpus vector with the highest similarity among the vectors;

[0023] The corpus vector with the highest similarity among the vectors is output as the recommended content.

[0024] Preferably, the user data is input into a pre-built large language model to obtain a selection strategy for each recall agent, including:

[0025] Inputting the user information into the large language model to obtain the weights of each recall agent;

[0026] Adjust the weight of each recall agent according to the preset diversity adjustment factor and the recall rate of each recall agent in the past;

[0027] The confidence of each recall agent is determined according to the adjusted recall agents as the selection strategy.

[0028] The embodiment of the present invention further provides a multi-agent based content recommendation device, the device comprising:

[0029] The content recommendation module is used to input the acquired user data into each pre-built recall agent to obtain the recommended content output by each recall agent;

[0030] A strategy generation module, used for inputting the user data into a pre-built large language model to obtain a selection strategy for each recall agent;

[0031] The fusion module is used to fuse the recommended contents output by each recall agent according to the selection strategy to obtain fused recommended contents.

[0032] Preferably, the device further comprises an interpretation module, configured to:

[0033] Inputting the user data and the fused recommended content into a preset interpretation agent to obtain recommended interpretation content;

[0034] The recommended explanation content and the fused recommended content are output as a recommendation result.

[0035] As a preferred solution, the fusion module is specifically used for:

[0036] Take the union of the recommendation results output by each recall agent to get the overall recommendation content;

[0037] Calculating the comprehensive score of each material in the overall recommended content according to the selection strategy;

[0038] The fused recommended content is selected according to the comprehensive scores of the various materials in the overall recommended content and preset screening rules.

[0039] Furthermore, the fusion module is specifically used for:

[0040] Taking the confidence in the selection strategy as a weight, weighted summing is performed on the recommendation scores of each material in the overall recommendation content in each recall agent to determine the recommendation score of each material in the overall recommendation content;

[0041] The recommendation scores of each item are input into a preset scoring calculation model, and the scores of each item in the overall recommended content in different dimensions are weighted to determine the comprehensive score of each item in the overall recommended content.

[0042] Preferably, the content recommendation module is specifically used for:

[0043] Input the acquired user data into each recall agent respectively to obtain the recommendation output of each recall agent;

[0044] Use a pre-trained vector model to convert each recommendation output into a vector representation in a high-dimensional space;

[0045] Each vector representation is used as a query statement, and compared with the material vectors in the preset vector database to determine the corpus vector with the highest similarity among the vectors;

[0046] The corpus vector with the highest similarity among the vectors is output as the recommended content.

[0047] Preferably, the strategy generation module is specifically used for:

[0048] Inputting the user information into the large language model to obtain the weights of each recall agent;

[0049] Adjust the weight of each recall agent according to the preset diversity adjustment factor and the recall rate of each recall agent in the past;

[0050] The confidence of each recall agent is determined according to the adjusted recall agents as the selection strategy.

[0051] An embodiment of the present invention also provides a terminal device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, a multi-agent-based content recommendation method as described in any of the above embodiments is implemented.

[0052] An embodiment of the present invention also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a multi-agent-based content recommendation method as described in any of the above embodiments.

[0053] An embodiment of the present invention further provides a computer program product, including a computer program / instruction, which implements the steps of the method described in any of the above embodiments when executed by a processor.

[0054] Compared with the prior art, the present invention provides a content recommendation method, device, equipment, medium and product based on multiple agents, which obtains the recommended content output by each recall agent by inputting the acquired user data into each pre-built recall agent respectively; inputs the user data into the pre-built large language model to obtain the selection strategy of each recall agent; and fuses the recommended content output by each recall agent according to the selection strategy to obtain the fused recommended content. The present application provides a more comprehensive output solution for recommended content by fusing the recommended content of multiple recall agents, improves the scope of application of scenarios, and performs content fusion based on the calculated selection strategy to improve the accuracy of content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flowchart of a multi-agent-based content recommendation method provided by an embodiment of the present invention;

[0056] Figure 2 is a schematic diagram of the structure of a recall agent provided by an embodiment of the present invention;

[0057] Figure 3 is a structural diagram of a multi-agent-based content recommendation device provided by an embodiment of the present invention;

[0058] Figure 4 It is a structural diagram of a terminal device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0059] 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.

[0060] It can be understood that the various embodiments described below can be combined or combined in a logical manner, and the embodiments of the present invention will not describe various combinations one by one.

[0061] Content-based recommendation methods rely on the analysis of item content and the matching of user preferences. This method usually involves extracting item features (such as text descriptions, keywords, tags, etc.) and building a user preference model based on the user's past behavior (such as ratings, browsing, purchases, etc.). New items that users may be interested in are identified and recommended by comparing the user preference model and item features.

[0062] The existing multi-agent-based content recommendation method uses a single model, which is difficult to apply to cold start and insufficient data richness scenarios, and the efficiency and accuracy of content recommendation are low.

[0063] In view of the technical problems existing in the above-mentioned prior art, this application proposes a content recommendation method based on multi-agent, see Figure 1 , is a flow chart of a multi-agent-based content recommendation method provided by an embodiment of the present invention, the method comprising the following steps:

[0064] Step S1, inputting the acquired user data into each pre-built recall agent respectively, and obtaining the recommended content output by each recall agent;

[0065] Step S2, inputting the user data into a pre-built large language model to obtain a selection strategy for each recall agent;

[0066] Step S3, fusing the recommended contents output by each recall agent according to the selection strategy to obtain fused recommended contents.

[0067] During the specific implementation of this embodiment, in order to avoid the problem of poor model fitness when a single recommendation model is sampled in the priority scheme to calculate the recommended content, this application completes the recommended content calculation process through multiple intelligent agents.

[0068] The acquired user data is input into each pre-built recall agent respectively to obtain the recommended content output by each recall agent.

[0069] This case adopts a multi-agent architecture, where each agent is supported by a large language model and is responsible for a specific task. According to the specific algorithm strategy and the input and output requirements defined for these strategies, each agent is configured with corresponding prompt words so that the large language model can efficiently solve specific tasks. Due to the differences in tasks, each agent may need to call external tools during execution to complete its task goals.

[0070] See also Figure 2 , is a schematic diagram of the structure of the recall agent provided by an embodiment of the present invention. The general architecture of the agent mainly includes the following modules:

[0071] Large language model: The core technology that supports the entire system, responsible for understanding and generating natural language, is the foundation of the intelligent agent.

[0072] Prompt word module: Provides contextual information of specific tasks to the large language model to ensure that the model can understand the task and generate corresponding solutions, mainly based on algorithm strategies and input and output requirements.

[0073] Tool modules: External tools that can be called by the agent to perform specific tasks and enhance the functionality and flexibility of the system, such as vector retrieval tools.

[0074] When building algorithm strategies, we will integrate user portrait information, context information, and historical user behavior data, including but not limited to browsing history, click history, and purchase history. Through in-depth analysis and mining of these data, we will build a user interest model to accurately reflect multi-dimensional information such as user interest preferences, consumption habits, and behavioral characteristics.

[0075] The process of constructing the interest model includes but is not limited to:

[0076] Feature engineering: Extract and transform user portrait information to obtain key information such as users’ basic attributes and preference tags.

[0077] Contextual information analysis: Consider the user’s current environment and situation, such as time, location, device, etc., to gain a deeper understanding of user behavior.

[0078] Behavioral data mining: Mining users’ historical behavioral data to explore their behavioral patterns, frequencies, and behavioral sequences, etc., in order to reveal their hidden interests and behavioral patterns.

[0079] Through the above steps, a user interest model is established, providing a solid foundation for multi-agent-based content recommendation and content customization.

[0080] Output data: Multi-agent based content recommendations and content suggestions based on user interest models. Recommended content covers products, services, information, etc. that users may be interested in to improve user satisfaction and engagement.

[0081] A plurality of different recall agents trained for different training data sets are used, and the acquired user data is input into the different recall agents to obtain the recommended contents output by the different recall agents.

[0082] It should be noted that the user data obtained in this case is the data for regular content recommendation, which is used to characterize the user's behavior characteristics and preference characteristics. By comparing the user preference model and item characteristics, new items that the user may be interested in are identified and recommended.

[0083] The multiple recall agent architecture is not just a simple collection of components, but a series of agents supported by advanced large language models, each of which is responsible for a specific part of the recommendation system (such as recall, ranking stage), or for a specific user group and content type. By leveraging the powerful learning and analysis capabilities of the large language model, the user data is input into the pre-built large language model to obtain the selection strategy of each recall agent.

[0084] It should be noted that the analysis of the selection strategy is based on the analysis of the data volume characteristics of the user data obtained, and then different recommendation scenarios are identified, for example, cold start scenarios are identified when the user profile and historical behavior data are missing or insufficient in the user data. Cold start scenarios of items are identified when the user profile and historical behavior data are rich. Behavior-based scenarios are identified when the user data summary user has rich historical behavior data, etc.

[0085] The selection strategy is the strategy for the degree of adoption of the content recommended by different recall agent models. That is, according to the characteristics of user data, the adoption probability of the content recommended by different recall agents is determined, and then the recommendation results are comprehensively determined from different recall agents according to the selection strategy.

[0086] After obtaining the selection strategies of different recall agents, the fusion recommendation content can be comprehensively selected from the recommendation contents of different recall agents according to the selection strategies.

[0087] In specific implementation, a ranking agent can be used alone to receive recommendation results from different recall agents, and these results can be efficiently processed in combination with a selection strategy to obtain fused recommendation content. Recommended content can be selected according to actual needs, and only the content with the highest score can be retained for display to ensure the quality of recommended content and user experience.

[0088] It should be noted that the selection strategy is specifically to determine the corresponding data scenario according to the user data characteristics, and then determine the recommended content of different recall agents to obtain the selection strategy. The selection strategy is used to characterize the degree of adoption of recommended content of different recall agents. The adoption degree is specifically characterized by ranking or using quantitative measurement standards, such as weight, credibility, confidence and other indicators.

[0089] When performing fusion processing, the recommended content with the highest degree of adoption is selected according to the selection strategy; or weighted fusion is performed according to the degree of adoption of different recall agents to obtain the fused recommendation result. It is also possible to select recall agents with weights within the preset range according to the weights of the degree of adoption of the recommended content of different recall agents and the preset trust range to determine the fused recommended content.

[0090] It can be seen from the above description that, according to the selection strategy, the implementation method of determining the fusion recommendation content is that those skilled in the art can infer a variety of feasible solutions based on the above description.

[0091] It should be noted that the process of obtaining user data in this case specifically includes:

[0092] Integrate context information based on query content, time and location of the received request information;

[0093] Extracting a user identifier from the request information, and querying user portrait data and historical behavior data according to the user identifier;

[0094] Using the integrated context information, the user portrait data and the historical behavior data as the user data;

[0095] Query user profiles specifically: query user preferences and other user profile information profiles based on user identifiers u .

[0096] Get user historical behavior data: Collect user historical behavior data action u , such as clicks, browsing behavior and other interactions.

[0097] Integrate context information: Combine the context information such as time, location, and device type in the request.

[0098] Each recall agent includes a user cold-start agent, an item cold-start agent, a behavior-based agent, a social-based agent, and a content-based agent obtained in advance based on different training data sets;

[0099] User cold start recall agent rule 1: When user profile and historical behavior data are missing or insufficient, the user cold start agent is selected to provide initial recommended content.

[0100] Item cold start recall agent rule 2: When the user profile and historical behavior data are rich, the item cold start agent can be selected to recommend newly launched or less widely interacted items to users.

[0101] Behavior-based recall agent rule 3: When the user has rich historical behavior data, select the behavior-based recall agent to generate recommended content by analyzing the user's behavior patterns and preferences.

[0102] Social recall agent rule 4: When the user's social relationships have a significant impact on the recommendation, a social-based recall agent is selected to make recommendations using the user's social network information (such as friend recommendations and social interactions).

[0103] Content-based recall agent rule 5: When recommendations need to be based on user interests and content similarity, select a content-based recall agent to generate recommended content by matching user interest tags and content features.

[0104] It should be noted that the above-mentioned user data and examples of each recall agent do not constitute a limitation on the solution. Technical personnel in this field can know the specific data type and quantity used by user data in combination with the prior art, and technical personnel in this field can adjust the number of recall agents according to actual needs and applicable scenarios.

[0105] This application provides a more comprehensive output solution for recommended content by integrating the recommended content of multiple recall agents, improving the adaptability of the solution to different scenarios and providing the accuracy of recommended content. The design of the recall agent can face scenarios such as cold start problems, data sparsity, and lack of diversity and novelty of recommended content.

[0106] In another embodiment provided by the present invention, the present application scheme further includes the following steps:

[0107] Step S4, inputting the user data and the fused recommended content into a preset interpretation agent to obtain recommended interpretation content;

[0108] Step S5: output the recommended explanation content and the fused recommended content as a recommendation result.

[0109] When this application is implemented, current recommendation explanations may include "based on your browsing history," "because you like similar brands," or "this is a popular choice this week," which are completely unable to meet the personalized needs of users.

[0110] In this proposal, the recommendation explanation agent integrates the basic attributes of each recommended material and combines the deep semantic understanding capabilities of the large language model to identify the key factors behind the recommendation, and uses the text generation capabilities of the large language model to generate personalized recommendation explanations for users.

[0111] The explanation framework can automatically generate explanations for recommendations based on the source and logic of the recommendations. Explanations are usually based on the following aspects:

[0112] Based on user behavior: Recommendations are based on specific user behavior, such as browsing and search history.

[0113] Based on item attributes: Recommended items have attributes that match the user's demonstrated interests.

[0114] Based on social relationships: Recommendations are based on the preferences of friends in the user's social network.

[0115] Time or location based: Recommendations are relevant to the current time or geographic location of the user.

[0116] Based on popularity: The recommended material is currently gaining widespread attention or is popular.

[0117] When training an explanation agent, it is necessary to use corresponding data for training according to the above-mentioned different explanation rules to obtain an explanation agent and construct an explanation relationship between user data and recommended content.

[0118] In specific implementation, user data and fused recommended content may be input into an explanatory agent, and explanatory content of the explanatory relationship between the user data and the recommended content may be output.

[0119] Input data: Combined with user profile information u , context information context and user's historical behavior data action u , generate a multi-agent based content recommendation explanation for the recommendation result recommend_result of the ranking agent. u ,context,action u ,recomm_result].

[0120] Output data: The generated multi-agent-based content recommendation explanation will be displayed together with the recommended content to enhance the user's understanding and trust in the recommendation system. explain ).

[0121] According to the above description, the recommendation explanation agent can automatically generate recommendation explanations based on the source and logic of the recommendation, making it easier for users to understand the recommended content.

[0122] It should be noted that the following are recommended explanations of the output:

[0123] Historical interaction analysis: Generate explanations closely related to the user's historical behavior based on the user's browsing, searching and purchasing records. For example: This camera is recommended to you because we noticed that you recently viewed a number of professional photography equipment.

[0124] Consider location and time: Provide explanations based on the user’s current location and possible seasons or special dates. For example: This cafe is recommended to you because it is located in your city and has good reviews.

[0125] Social Preference Mapping: If applicable, introduce the user's social preferences as a basis for recommendation. For example: Your friend Anna recently visited this restaurant and gave it a five-star rating.

[0126] Context-adaptive explanations: Consider the specific context or upcoming events facing the user and generate relevant explanations for recommendations. For example, since you are going to attend an outdoor event next week, this waterproof action camera would be a good choice for you.

[0127] Multi-agent based content recommendation explanations aim to provide clear and intuitive explanations for the ranked recommendation results, helping users understand why these items are recommended to them. This can not only improve user satisfaction, but also enhance the transparency and credibility of the recommendation system.

[0128] In another embodiment provided by the present invention, the step S3 specifically includes the following steps:

[0129] Receive the recommendation results from different recall agents and perform a deduplication operation, that is, find the union of the recommendation results output by different recall agents to obtain the overall recommendation content, ensuring that there are no duplicate items in the recommendation list. Assume that the recommendation result set obtained from different recall agents is R1, R2, ..., R n , then the overall recommended content after deduplication is R unique We can remove duplicates by taking the union of the sets and get: unique =R1∪R2∪...∪R n .

[0130] For the overall recommended content, the comprehensive score of each material is calculated based on the scoring calculation model and the selection strategy of the recall agent combined with its source;

[0131] For fusion processing, the importance or confidence of the results of each recall agent and the similarity of each recalled material can be considered. Then the fusion processing needs to integrate the scores of different recall agents on the same material.

[0132] It should be noted that there are duplicate materials in the recommended contents output by different recall agents. The more recall agents recommend materials, the higher the confidence of the recommended recall agents and the higher the score of the materials.

[0133] Sort all materials according to their ratings to get the final fused recommendation list R merge . The list is arranged from highest to lowest overall score for the material.

[0134] The fused recommended content is selected according to a preset screening rule.

[0135] For example, the recommendation list can be truncated according to actual needs, retaining only the top N materials with the highest scores for display, or selecting materials whose scores are within a preset range.

[0136] By scoring the recommendation results of each recall agent and determining the best material based on the score, the quality of the recommended content is ensured. The recommendation results are accurately sorted according to the user's personalized needs and interest model to ensure that the most relevant content is displayed first.

[0137] In another embodiment provided by the present invention, the process of determining the score of each material specifically includes the following steps:

[0138] Taking the confidence in the selection strategy as a weight, weighted summing is performed on the recommendation scores of each material in the overall recommendation content in each recall agent;

[0139] Consider the importance or confidence of each recall agent’s result and the similarity of each recalled item. Assume that each item j In R i There is a score or weight in each item, which can be expressed as score(item j ,R i )=sim j , then the fusion processing needs to consider how to integrate the scores of different recall agents on the same material.

[0140] For each material item j The result set R after deduplication unuque The comprehensive score in each R i The scores in the weighted formula are used to obtain the recommendation scores of each material in the overall recommendation content.

[0141] Among them, conf i is the recall agent R i The confidence of item indicates its relative importance in the result fusion. j Not in R i In, sim j =0, or a base score can be set to reflect the difference in recall capabilities of different recall agents.

[0142] In addition to sim-based j In addition to the basic sorting, the sorting process can also consider factors in multiple dimensions, such as the matching degree of user's personal interests, the novelty of the material, the interaction history between the user and the material, etc.

[0143] Input each recommendation score into the preset scoring calculation model, perform weighted calculation on the scores of each material in different dimensions, and determine the comprehensive score of each material in the overall recommendation content. j =

[0144] α1×S(item j )+α2×Novelty(item j )+...+α k ×Interaction(item j ,action u ).

[0145] Among them, α1, α2, ..., α k It is the weight of each factor score, which needs to be adjusted through system optimization and user feedback. Novelty represents the novelty of the material, Interaction (item j ,action u ) The degree of match between the user’s interaction history and the item.

[0146] Finally, all materials are sorted according to the comprehensive scores of each material to obtain the final fused recommendation list R merge This list is arranged from high to low according to the comprehensive score of the items, aiming to provide a recommendation result that integrates the opinions of multiple recall agents.

[0147] The comprehensive score of each material is calculated based on a multi-dimensional rating calculation model, and materials are recommended based on the comprehensive score, maintaining the diversity and novelty of the recommendation list, avoiding content uniformity, and providing users with more choices and new discoveries.

[0148] In another embodiment provided by the present invention, the step S1 specifically includes the following steps:

[0149] The acquired user data is input into each recall agent respectively to obtain the recommendation output of each recall agent.

[0150] User profile information u A combination of context information and the user's historical behavior data action u These data will be used as components of the prompt words for subsequent processing.

[0151] input=[profile u ,context,action u ];

[0152] Output data: Multi-agent based content recommendations and content suggestions based on user interest models. Recommended content covers products, services, information, etc. that users may be interested in to improve user satisfaction and engagement.

[0153] recomm_info=LLM(input,rule recall );

[0154] The recommended output data is then vectorized, and the recommended content is converted into a vector representation in a high-dimensional space through advanced deep learning models, such as pre-trained vector models (Embedding Model, EM) m3e or gte. This step uses the powerful capabilities of the vector model to encode the natural language query sentence into a dense vector. This vector is not only a form of representation, but more importantly, it captures the rich semantic features of the sentence, providing a solid foundation for subsequent similarity matching and content retrieval.

[0155] The vector is represented as embedding = EM (recomm_info).

[0156] Each vector representation is used as a query statement and compared with the material vectors in the preset vector database to determine the corpus vector with the highest similarity among the vectors.

[0157] The embedding vector of the query statement is compared with the embedding vector of the material library stored in the vector database (VectorDB). This comparison process relies on the similarity calculation between vectors, aiming to find the item vector that is closest to the user's query statement vector. Each item in the material library is preprocessed by the same or compatible deep learning model to ensure that its embedding vector can accurately represent the core features and attributes of the item.

[0158] Through this method, the intelligent agent can quickly and accurately filter out items that are highly relevant to compound query statements from a large-scale material library and provide users with personalized recommendation results.

[0159] sim=VectorDB(database, embedding, topk).

[0160] Among them, sim∈R k It is a k-dimensional array output by the vector database, i.e., the recall result, where each element corresponds to an item. Database refers to the pre-generated embedding of the entire item library. Topk selects the k most matching items.

[0161] Based on the user interest model, compound query statements are generated and converted into vector representations in high-dimensional space through a deep learning model for subsequent content retrieval and matching. The generation process of compound query statements integrates the user's multi-dimensional interests and needs, which can improve the relevance and accuracy of recommended content.

[0162] In another embodiment provided by the present invention, the step S2 specifically includes:

[0163] Inputting the user information into the large language model to obtain the weights of each recall agent;

[0164] Input data: User profile u , User historical behavior action u And context information context is integrated as input data to construct the prompt word.

[0165] input=[profile u ,action u ,context];

[0166] rules=[rule1,rule2,rule3,rule4,rule5];

[0167] Output data: Define the data format and content of the large language model output, and generate a weight array to represent the probability of each recall agent being selected.

[0168] Output weight weight = LLM (input, rules).

[0169] Among them, weight∈R n It is an n-dimensional vector output by the large language model, where each element corresponds to the weight of a recall agent.

[0170] The weight of each recall agent is adjusted according to the preset diversity adjustment factor and the recall rate of each recall agent in the past.

[0171] Introduce diversity adjustment factors and adjust weights

[0172] Among them, α is a tuning parameter used to control the influence of diversity on weight. i is the diversity performance of the i-th recall agent.

[0173] Consider the adjustment factor of historical performance, and the adjusted weight is weight i ′=weight i ×(1+β×recall i ).

[0174] Among them, β is the adjustment parameter, recall i represents the recall rate of the i-th recall agent in the past.

[0175] The recall agent is selected based on the adjusted weights. The selection probability can still be expressed using the weight-normalized value. The weights of each recall agent after adjustment

[0176] Determine the confidence of each recall agent based on the adjusted recall agents

[0177] For example, for a cold-start user, the recommendation system can obtain the user’s social friends, recall the agent’s initial weight weight = [0.5, 0.1, 0.1, 0.3, 0.1], diversity index diversity = [0.1, 0.5, 0.3, 0.3, 0.1], recall index recall = [0.2, 0.1, 0.15, 0.2, 0.25], and adjust parameters α = 2, β = 1.

[0178] Calculate the adjusted weight: weight′=[1.01665,0.16,0.1775,0.5475,0.20833].

[0179] Calculate confidence: conf = [0.342, 0.145, 0.148, 0.214, 0.152].

[0180] Introducing recall rate as an adjustment factor can help the system adjust the weight of the recall agent more carefully based on historical performance, thereby achieving a more accurate and efficient recommendation strategy.

[0181] See also Figure 3 , is a schematic diagram of a structure of a multi-agent-based content recommendation device provided by an embodiment of the present invention, the device comprising:

[0182] The content recommendation module is used to input the acquired user data into each pre-built recall agent to obtain the recommended content output by each recall agent;

[0183] A strategy generation module, used for inputting the user data into a pre-built large language model to obtain a selection strategy for each recall agent;

[0184] The fusion module is used to fuse the recommended contents output by each recall agent according to the selection strategy to obtain fused recommended contents.

[0185] The multi-agent based content recommendation device provided in this embodiment can execute all the steps and functions of the multi-agent based content recommendation method provided in any of the above embodiments, and the specific functions of the device will not be described in detail here.

[0186] See also Figure 4, is a schematic diagram of the structure of a terminal device provided in an embodiment of the present invention. The terminal device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a multi-agent-based content recommendation program. When the processor executes the computer program, the steps in the above-mentioned embodiments of the multi-agent-based content recommendation method are implemented, such as Figure 1 Alternatively, the processor implements the functions of each module in the above-mentioned device embodiments when executing the computer program.

[0187] Exemplarily, the computer program 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 that can perform specific functions, and the instruction segments are used to describe the execution process of the computer program in the multi-agent-based content recommendation device. For example, the computer program may be divided into several modules, and the specific functions of each module have been described in detail in the multi-agent-based content recommendation method provided in any of the above embodiments, and the specific functions of the device will not be repeated here.

[0188] The multi-agent-based content recommendation device may be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The multi-agent-based content recommendation device may include, but is not limited to, a processor and a memory. Those skilled in the art may understand that the schematic diagram is merely an example of a multi-agent-based content recommendation device and does not constitute a limitation on a multi-agent-based content recommendation device. It may include more or fewer components than shown in the diagram, or a combination of certain components, or different components. For example, the multi-agent-based content recommendation device may also include input and output devices, network access devices, buses, etc.

[0189] 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 multi-agent-based content recommendation device, and uses various interfaces and lines to connect the various parts of the multi-agent-based content recommendation device.

[0190] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the multi-agent-based content recommendation device 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, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, 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 Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0191] Wherein, if the module integrated in the multi-agent-based content recommendation device 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, etc. 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, etc.

[0192] An embodiment of the present invention further provides a computer program product, including a computer program / instruction, which implements the steps of any method described in the above embodiments when executed by a processor.

[0193] It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A content recommendation method based on multi-agent, characterized in that: The method comprises: Input the acquired user data into each pre-built recall agent respectively, and obtain the recommended content output by each recall agent; Inputting the user data into a pre-built large language model to obtain a selection strategy for each recall agent; The recommendation contents output by each recall agent are fused according to the selection strategy to obtain fused recommendation contents.

2. The multi-agent based content recommendation method according to claim 1, characterized in that: The method further comprises: Inputting the user data and the fused recommended content into a preset interpretation agent to obtain recommended interpretation content; The recommended explanation content and the fused recommended content are output as a recommendation result.

3. The multi-agent based content recommendation method according to claim 1, characterized in that: The step of fusing the recommended contents output by each recall agent according to the selection strategy to obtain the fused recommended contents includes: Take the union of the recommendation results output by each recall agent to get the overall recommendation content; Calculating the comprehensive score of each material in the overall recommended content according to the selection strategy; The fused recommended content is selected according to the comprehensive scores of the various materials in the overall recommended content and preset screening rules.

4. The multi-agent based content recommendation method according to claim 3, characterized in that: Calculating the comprehensive score of each material in the overall recommended content according to the selection strategy includes: Taking the confidence in the selection strategy as a weight, weighted summing is performed on the recommendation scores of each material in the overall recommendation content in each recall agent to determine the recommendation score of each material in the overall recommendation content; The recommendation scores of each item are input into a preset scoring calculation model, and the scores of each item in the overall recommended content in different dimensions are weighted to determine the comprehensive score of each item in the overall recommended content.

5. The multi-agent based content recommendation method according to claim 1, characterized in that: The acquired user data is input into each pre-built recall agent to obtain the recommended content output by each recall agent, including: Input the acquired user data into each recall agent respectively to obtain the recommendation output of each recall agent; Use a pre-trained vector model to convert each recommendation output into a vector representation in a high-dimensional space; Each vector representation is used as a query statement, and compared with the material vectors in the preset vector database to determine the corpus vector with the highest similarity among the vectors; The corpus vector with the highest similarity among the vectors is output as the recommended content.

6. The multi-agent based content recommendation method according to claim 1, characterized in that: The user data is input into a pre-built large language model to obtain the selection strategies of each recall agent, including: Inputting the user information into the large language model to obtain the weights of each recall agent; Adjust the weight of each recall agent according to the preset diversity adjustment factor and the recall rate of each recall agent in the past; The confidence of each recall agent is determined according to the adjusted recall agents as the selection strategy.

7. A content recommendation device based on multi-agent, characterized in that: The device comprises: The content recommendation module is used to input the acquired user data into each pre-built recall agent to obtain the recommended content output by each recall agent; A strategy generation module, used for inputting the user data into a pre-built large language model to obtain a selection strategy for each recall agent; The fusion module is used to fuse the recommended contents output by each recall agent according to the selection strategy to obtain fused recommended contents.

8. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the multi-agent-based content recommendation method as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the multi-agent-based content recommendation method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of any method described in claims 1 to 6 are implemented.

Images