A multi-agent collaborative recommendation system based on LLM

Abstract

The application discloses a multi-agent system recommendation system based on LLM, which comprises a knowledge infrastructure partition autonomous construction module, an LLM-driven user deep understanding module and a multi-agent collaborative retrieval and recommendation module. The system realizes intelligent content partition storage and knowledge organization through a self-developed classification model, performs semantic reconstruction and deep intention reasoning on a user behavior sequence based on LLM, and then completes retrieval and sorting tasks in parallel by multiple special agents, and dynamically fuses the results through a gating network. The application can improve the diversity and explorability of recommended content, accurately understand the deep needs of users, reduce the computing cost and improve the reasoning speed, effectively enhance the generalization ability of the system, solve the problems of information cocoon, lack of semantic reasoning and low efficiency of the traditional recommendation system, realize intelligent recommendation with accuracy, diversity and high efficiency, and is suitable for various personalized recommendation scenes such as news and e-commerce.

Description

Technical Field

[0001] This invention relates to the field of recommender system technology, and in particular to a multi-agent collaborative recommender system based on LLM. Background Technology

[0002] With the explosive growth of internet information, recommender systems have become a key technology for alleviating information overload and improving user experience. Existing recommender systems are mainly based on collaborative filtering, content filtering, and hybrid models of these, and are gradually incorporating technologies such as deep learning and knowledge graphs.

[0003] The closest existing technologies mainly include: 1. Deep learning-based sequence recommendation systems: These systems utilize recurrent neural networks or Transformer models to process user historical behavior sequences and predict the next interactive item. While these methods can capture temporal patterns, they essentially still rely on historical co-occurrence statistics and lack semantic understanding of deep user intent. 2. Knowledge graph-based recommendation systems: These systems construct entity relationship networks between items and use graph neural networks for reasoning to enhance the diversity and interpretability of recommendations. However, these methods typically rely on manually constructed or limited-source knowledge graphs, making it difficult to dynamically understand complex and abstract user needs. 3. Pre-trained language model-based recommendation systems: These systems attempt to use large language models to process textual content, such as item descriptions or user reviews. However, directly using general-purpose large language models for recommendations often faces problems such as high computational costs, large inference latency, and lack of deep integration with efficient retrieval and ranking architectures.

[0004] The common drawbacks of existing technologies are: First, over-reliance on users' historical preference data can easily lead to "information cocoons," resulting in highly homogenized recommendation results and limiting users' exploration; second, a lack of deep reasoning ability regarding the underlying motivations and contextual semantics of user behavior leads to insufficient generalization ability when handling new items or complex queries; and finally, the direct application of large models results in low system efficiency, making it difficult to meet the response speed requirements of real-time recommendation scenarios. Summary of the Invention

[0005] The purpose of this invention is to solve the problems mentioned in the background art by proposing a multi-agent collaborative recommendation system based on LLM.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A multi-agent cooperative recommendation system based on LLM includes: The knowledge infrastructure partitioned autonomy construction module is used to realize multimodal content processing, intelligent partitioning, knowledge vectorization, partitioned knowledge graph construction and partitioned retrieval enhancement services, while building a user behavior data pipeline to collect and store user behavior data in real time and in layers. The LLM-driven user deep understanding module is used to acquire and integrate multi-source user data, semantically reconstruct user behavior sequences, mine graph relationships in behavior sequences, perform deep reasoning through LLM to gain insights into users' deep needs, and generate accurate search suggestions. The multi-agent collaborative retrieval and recommendation module includes multiple dedicated agents and a gating network. The dedicated agents include at least a vector semantic retrieval agent and a graph reasoning agent, which are used to complete multimodal content retrieval based on user needs. The gating network is used to weight and fuse the retrieval results to form a candidate set, and then sort them through coarse and fine ranking agents to generate the final recommendation list.

[0007] As a further aspect of the present invention: the knowledge infrastructure partitioned autonomous construction module includes: The multimodal content understanding unit uses the self-developed Dakon-OCR model to process text and image-based multimodal content, extracting titles, body text, and metadata, and generating structured semantic information. The intelligent content partitioning unit adopts the Dakon-Kangxi classification model based on the Chinese-RoBERTa-wwm-ext-large architecture. It achieves accurate content classification by analyzing the first 150 characters of the content, with a classification accuracy rate of >99%. The partition knowledge vectorization unit converts the semantic content of each partition into a high-dimensional vector and stores it in the corresponding partition in Elasticsearch; The partitioned knowledge graph construction unit uses Dakon-Graph technology to extract entity relationships within each partition, constructs partitioned subgraphs, and stores them in NebulaGraph. The partitioned retrieval enhancement unit provides a fusion service of semantic retrieval and graph retrieval within a partition through the Dakon-RAG framework.

[0008] As a further aspect of the present invention: the user behavior data pipeline includes: The real-time data collection unit collects user behavior data such as clicks and reads in real time through a Kafka message queue. The tiered storage unit routes the collected user behavior data to the corresponding partition directory of the HDFS data lake, and stores structured user profile data and unstructured behavior sequence description data in a tiered manner.

[0009] As a further aspect of the present invention: the LLM-driven user deep understanding module specifically performs the following steps: S1: Obtain the user behavior sequence from HDFS and call the Dakon-RAG service to retrieve detailed information about the corresponding content; S2: Reconstruct the raw user behavior logs into a semantic description in natural language form; S3: Analyze hidden relationships in behavioral sequences using Dakon-Graph to form a user interest network; S4: Input semantic descriptions and interest networks into the LLM, and let the LLM infer and output the user's deep needs; S5: LLM generates precise search suggestions based on users' deep needs.

[0010] As a further aspect of the present invention: the retrieval and ranking process of the multi-agent collaborative retrieval and recommendation module includes: Step 1: The vector semantic retrieval agent performs similarity matching retrieval in the corresponding ElasticSearch partition based on the user's requirement embedding; the graph reasoning agent mines entity paths related to the user's interests through NebulaGraph and completes the retrieval. Step 2: The gating network dynamically assigns weights to the retrieval results of the two agents based on the user context and merges them to form a candidate set; Step 3: Multiple coarse-ranking agents filter the candidate set in parallel based on popularity and relevance indicators, and output the Top-K content; Step 4: Multiple ranking agents score the Top-K content from the dimensions of relevance, diversity, and novelty. The gating network merges the scoring results and re-ranks them to generate the final recommendation list.

[0011] As a further aspect of the present invention, the weight allocation strategy of the gating network can be dynamically adjusted, giving higher weights to the graph reasoning agent for exploratory users.

[0012] As a further solution of the present invention, alternative implementations of the intelligent content partitioning unit include fine-tuning using lightweight pre-trained models such as ALBERT and TinyBERT, or using a combination of rule-based initial partitioning and LLM optimization.

[0013] As a further solution of the present invention: alternative implementations of the LLM-driven user deep understanding module include using multiple dedicated small models to process interest patterns and evolutionary analysis dimensions respectively, and then fusing the results through ensemble learning; or using a rule engine combined with a knowledge graph for reasoning.

[0014] Compared with existing technologies, the advantages of this invention are: 1. Breaking through the limitations of information cocoons, through the autonomous design of knowledge infrastructure and collaborative retrieval by multiple agents, combined with the proactive exploration of diverse agents, the diversity and cross-domain exploration of recommended content are effectively improved, and users' horizons are broadened; 2. Achieve accurate matching of deep needs: Based on semantic reconstruction and deep reasoning of LLM, combined with relationship mining of knowledge graph, it can accurately understand deep intentions from user behavior sequences, surpass the limitations of traditional similarity matching, and improve recommendation accuracy; 3. Improve system performance: Reduce computational costs through lightweight self-developed models (Dakon-Kangxi, Dakon-OCR, etc.) and partitioned architecture design; improve inference speed and real-time performance by combining multi-agent parallel work with dynamic fusion of gating networks to meet the needs of practical application scenarios. 4. Enhanced generalization ability: The partitioned knowledge organization and multi-agent collaborative strategy enable the system to effectively handle scenarios such as new items and complex user queries, thereby improving the system's adaptability and robustness. Attached Figure Description

[0015] Figure 1 This is a diagram of the overall system architecture of the present invention.

[0016] Figure 2 This is a flowchart illustrating the construction process of the knowledge infrastructure for this invention.

[0017] Figure 3 This is a flowchart for user-level in-depth understanding of the present invention.

[0018] Figure 4 This is a flowchart of the multi-agent collaborative retrieval and recommendation process of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] A multi-agent collaborative recommendation system based on LLM, such as Figure 1 As shown, this system presents a layered collaborative architecture, specifically including: Input layer: Receives multimodal content (text, images, videos) and user behavior data (clicks, views, comments, etc.).

[0021] Processing layer: This is the core computing layer, which includes two core processes: knowledge infrastructure construction and LLM-driven deep user understanding.

[0022] The agent layer consists of multiple specialized agents (such as semantic retrieval, graph reasoning, and diversity exploration agents), which are computational entities that perform collaborative retrieval and recommendation.

[0023] Control layer: Composed of MOA (Multi-Agent Orchestration) gating network, responsible for dynamic arbitration, weight allocation and result fusion of the agent layer.

[0024] Data storage layer: A hybrid storage solution is adopted, including ElasticSearch vector database, NebulaGraph graph database, HDFS data lake and MySQL relational database, to provide data support for the upper layer.

[0025] Output layer: Generates and displays the final personalized recommendation list.

[0026] Taking news recommendations as an example, suppose user User123 has read multiple news articles about "quantum computing" and "the application of AI in biological sciences" in the past 30 days. Figures 1 to 4 This system needs to generate a personalized recommendation list based on this user behavior.

[0027] Phase 1: Construction of Autonomous Regions for Knowledge Infrastructure This phase involves processing and organizing news content, as well as building a user behavior data pipeline. The specific steps are as follows: (1) Multimodal content understanding: The Dakon-OCR model is used to process the news articles to be recommended. For example, the article titled "New Breakthrough in Quantum Computer" is parsed to extract metadata such as title, body text, and publication time to generate structured semantic information. (2) Intelligent content partitioning: The Dakon-Kangxi classification model is called to analyze the content of the above article. By parsing the first 150 characters of the article, it is determined that the article belongs to the "technology" field and is assigned to the "technology" partition. (3) Knowledge vectorization by partition: Within the “Technology” partition, the semantic information of the article is converted into a high-dimensional vector and stored in the “Technology” partition of ElasticSearch; (4) Construction of partitioned knowledge graph: Extract entity relationships in the article through Dakon-Graph technology, such as the relationship between “quantum computing” and “AI”, the potential relationship between “quantum computing” and “biological science”, etc., and construct a sub-graph of the “technology” domain, which is stored in NebulaGraph; (5) Enhanced search service for the “Technology” partition: Deploy enhanced search service for the “Technology” partition based on the Dakon-RAG framework to ensure that subsequent search requests for the user are only made within the “Technology” partition and avoid cross-domain interference; (6) User behavior data pipeline operation: The click and reading behavior data of User123 are collected in real time through the Kafka message queue and routed to the "Technology" partition directory of the HDFS data lake; the data is processed in layers, and the structured user profile data (such as user ID and interest tag preliminary classification) and the unstructured behavior sequence description data (such as reading time and article title) are stored separately to prepare for subsequent LLM understanding.

[0028] Phase Two: In-depth User Understanding of LLM Drivers This phase involves uncovering the deeper needs of User123. The specific steps are as follows: (1) Acquisition and fusion of multi-source data: Obtain User123’s behavioral sequence data from HDFS (such as “User123, Monday, click, new breakthrough in quantum computing” and “User123, Tuesday, read, AI for protein folding prediction”), and call Dakon-RAG service to obtain detailed information of these two articles (such as tags “quantum computing”, “AI”, and “bioscience”). (2) Semantic reconstruction of behavior sequence: The original behavior log is reconstructed into a natural language description: "User123 clicked to read 'New Breakthrough in Quantum Computers' on Monday and read 'AI for Protein Folding Prediction' on Tuesday"; (3) Graph relationship mining: By analyzing the above behavior sequence through Dakon-Graph, it was found that User123's interests gradually evolved from "quantum computing" to "AI applications in biological sciences", forming an interest network containing "quantum computing-AI-biological sciences"; (4) LLM Deep Reasoning and Demand Insight: Input semantic description and interest network into LLM, and output the user's deep needs after LLM reasoning: "The user may be a scientific researcher who wants to track the progress of the cross-application of quantum computing and AI in life sciences." (5) Precise prompt word generation: LLM generates search prompt words based on deep needs: "Looking for recent cutting-edge research on the cross-application of quantum computing and AI in life sciences and materials discovery, requiring novel and in-depth content."

[0029] Phase 3: Multi-agent collaborative retrieval and recommendation This stage involves retrieving and sorting candidate content to generate the final recommendation list. The specific steps are as follows: (1) Multimodal content retrieval: ① The vector semantic retrieval agent performs similarity matching in the "Science and Technology" section of ElasticSearch based on the embedding of the search prompt words, and retrieves articles such as "Quantum Computing and AI-Driven Drug Discovery" and "AI + Quantum Computing Empowering Life Sciences"; ② The graph reasoning agent mines the association paths such as "Quantum Computing-Drug Discovery" and "AI-Protein Folding" through the "Science and Technology" subgraph of NebulaGraph, and retrieves articles such as "Application of Quantum Computing in Drug Development"; ③ The gating network determines that User123 is an exploratory user, assigns 60% weight to the retrieval results of the graph reasoning agent, assigns 40% weight to the vector semantic retrieval agent, and merges them to form a candidate set containing 10 articles; (2) Intelligent ranking and recommendation: ① Coarse ranking: Three coarse ranking agents filter the candidate set in parallel based on popularity, relevance and timeliness indicators, and output the top 5 content; ② Fine ranking and re-ranking: Three fine ranking agents score the top 5 content from the dimensions of relevance, diversity and novelty. The diversity agent introduces the article "Cross-application of quantum computing in the energy field" to enhance diversity; the gating network merges the scoring results to complete the re-ranking; ③ Final recommendation generation: Output the ranked recommendation list, including articles such as "The latest progress of quantum computing and AI in biomedicine", "AI + quantum computing empowers life sciences" and "Application of quantum computing in drug development".

[0030] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-agent collaborative recommendation system based on LLM, characterized in that, include: The knowledge infrastructure partitioned autonomy construction module is used to realize multimodal content processing, intelligent partitioning, knowledge vectorization, partitioned knowledge graph construction and partitioned retrieval enhancement services, while building a user behavior data pipeline to collect and store user behavior data in real time and in layers. The LLM-driven user deep understanding module is used to acquire and integrate multi-source user data, semantically reconstruct user behavior sequences, mine graph relationships in behavior sequences, perform deep reasoning through LLM to gain insights into users' deep needs, and generate accurate search suggestions. The multi-agent collaborative retrieval and recommendation module includes multiple dedicated agents and a gating network. The dedicated agents include at least a vector semantic retrieval agent and a graph reasoning agent, which are used to complete multimodal content retrieval based on user needs. The gating network is used to weight and fuse the retrieval results to form a candidate set, and then sort them through coarse and fine ranking agents to generate the final recommendation list.

2. The multi-agent collaborative recommendation system based on LLM according to claim 1, characterized in that, The knowledge infrastructure partitioned autonomy construction module includes: The multimodal content understanding unit uses the self-developed Dakon-OCR model to process text and image-based multimodal content, extracting titles, body text, and metadata, and generating structured semantic information. The intelligent content partitioning unit adopts the Dakon-Kangxi classification model based on the Chinese-RoBERTa-wwm-ext-large architecture. It achieves accurate content classification by analyzing the first 150 characters of the content, with a classification accuracy rate of >99%. The partition knowledge vectorization unit converts the semantic content of each partition into a high-dimensional vector and stores it in the corresponding partition in Elasticsearch; The partitioned knowledge graph construction unit uses Dakon-Graph technology to extract entity relationships within each partition, constructs partitioned subgraphs, and stores them in NebulaGraph. The partitioned retrieval enhancement unit provides a fusion service of semantic retrieval and graph retrieval within a partition through the Dakon-RAG framework.

3. The multi-agent collaborative recommendation system based on LLM according to claim 2, characterized in that, The user behavior data pipeline includes: The real-time data collection unit collects user behavior data such as clicks and reads in real time through a Kafka message queue. The tiered storage unit routes the collected user behavior data to the corresponding partition directory of the HDFS data lake, and stores structured user profile data and unstructured behavior sequence description data in a tiered manner.

4. The multi-agent collaborative recommendation system based on LLM according to claim 3, characterized in that, The LLM-driven user deep understanding module specifically performs the following steps: S1: Obtain the user behavior sequence from HDFS and call the Dakon-RAG service to get the detailed information of the corresponding content; S2: Reconstruct the raw user behavior logs into a semantic description in natural language form; S3: Analyze hidden relationships in behavioral sequences using Dakon-Graph to form a user interest network; S4: Input semantic descriptions and interest networks into the LLM, and let the LLM infer and output the user's deep needs; S5: LLM generates precise search suggestions based on users' deep needs.

5. The multi-agent collaborative recommendation system based on LLM according to claim 4, characterized in that, The retrieval and ranking process of the multi-agent collaborative retrieval and recommendation module includes: Step 1: The vector semantic retrieval agent performs similarity matching retrieval in the corresponding ElasticSearch partition based on the user's requirement embedding; the graph reasoning agent mines entity paths related to the user's interests through NebulaGraph and completes the retrieval. Step 2: The gating network dynamically assigns weights to the retrieval results of the two agents based on the user context and merges them to form a candidate set; Step 3: Multiple coarse-ranking agents filter the candidate set in parallel based on popularity and relevance indicators, and output the Top-K content; Step 4: Multiple ranking agents score the Top-K content from the dimensions of relevance, diversity, and novelty. The gating network merges the scoring results and re-ranks them to generate the final recommendation list.

6. The multi-agent collaborative recommendation system based on LLM according to claim 5, characterized in that, The weight allocation strategy of the gating network can be dynamically adjusted, giving higher weights to the graph reasoning agent for exploratory users.

7. The multi-agent cooperative recommendation system based on LLM according to claim 6, characterized in that, Alternative implementations of the intelligent content partitioning unit include fine-tuning using lightweight pre-trained models such as ALBERT and TinyBERT, or using a combination of rule-based initial partitioning and LLM optimization.

8. The multi-agent cooperative recommendation system based on LLM according to claim 7, characterized in that, Alternative implementations of the LLM-driven user deep understanding module include using multiple dedicated small models to process interest patterns and evolutionary analysis dimensions respectively, and then fusing the results through ensemble learning; or using a rule engine combined with a knowledge graph for reasoning.

Images