Intelligent recommendation system and algorithm based on knowledge graph

Abstract

The invention discloses an intelligent recommendation system and algorithm based on a knowledge graph, and relates to the technical field of information recommendation. The intelligent recommendation system based on the knowledge graph and the dynamic updating method thereof are achieved mainly by taking dynamic updating of the knowledge graph, multi-modal information fusion, cold start problem solving and recommendation algorithm optimization as the core, and the purpose is to establish and maintain the association relation graph between the user and the article and combine real-time data analysis to achieve the intelligent recommendation system based on the knowledge graph and the dynamic updating method of the intelligent recommendation system based on the knowledge graph. And the recommendation algorithm is continuously optimized to improve the accuracy and timeliness of recommendation.

Description

Technical Field

[0001] The present invention relates to the field of information recommendation technology, and in particular to an intelligent recommendation system and algorithm based on knowledge graphs. Background Art

[0002] Against the backdrop of accelerating digital transformation, the amount of information is growing exponentially, and users face a serious dilemma in information sifting. Recommendation systems, as a core technology solution for addressing information overload, have been widely used in e-commerce platforms, social networks, online education, and other fields, effectively improving information acquisition efficiency and user experience. However, traditional recommendation systems have many limitations:

[0003] (1) Insufficient recommendation accuracy and personalization: Traditional recommendation systems rely heavily on user historical behavior data and collaborative filtering algorithms, which can only capture surface information about user behavior and are unable to deeply understand users’ complex and ever-changing interests. For example, in e-commerce recommendations, recommendations are made based solely on items that users have purchased, ignoring the complex semantic relationships between items. This leads to significant deviations between recommendation results and users’ underlying preferences, making it difficult to accurately capture users’ personalized needs, thus reducing the practicality of the recommendation system and user satisfaction.

[0004] (2) Significant cold start challenges: For new users and new items, due to the lack of historical interaction data, traditional recommendation models find it difficult to establish accurate user profiles and item representations, resulting in a significant decline in recommendation quality. For example, for newly launched online courses, the system cannot determine its potential audience based on past user behavior, making it difficult to promote the course and for users to discover new products or content that meet their needs.

[0005] (3) Inefficient use of multimodal information: In actual application scenarios, user behavior data and item information often exist in multimodal forms such as text, images, and videos. Traditional recommendation systems find it difficult to effectively integrate and utilize these heterogeneous data sources. However, traditional recommendation systems mainly use single-modal data (such as user ratings and click behavior data) for recommendations and are unable to fully capture user interests. For example, in video recommendations, only user viewing history is considered, and multimodal information such as images and audio in the video is not fully utilized, resulting in poor recommendation results.

[0006] With the rise of knowledge graph technology, which effectively organizes entities, attributes, and relationships in a graph structure, providing rich semantic information and contextual associations for recommendation systems, knowledge graph-based recommendation methods offer a new technical path to address the limitations of traditional recommendation systems. While significant research has been achieved in this field, many challenges remain in key technical areas such as knowledge representation, graph construction, and reasoning applications. Summary of the Invention

[0007] The purpose of the present invention is to address the above-mentioned shortcomings and propose a knowledge graph-based intelligent recommendation system and its dynamic update method, which continuously optimizes the recommendation algorithm by constructing and maintaining the association relationship graph between users and items and combining real-time data analysis.

[0008] The present invention specifically adopts the following technical solutions:

[0009] Intelligent recommendation system based on knowledge graph, including

[0010] (1) Dynamic update mechanism of knowledge graph, including:

[0011] First, build an event-driven dynamic update framework

[0012] Design an event-driven knowledge graph update framework that triggers knowledge graph updates by monitoring user behavior and external environment changes. Use stream processing to capture user behavior data in real time and use it as input for knowledge graph updates.

[0013] Secondly, the incremental knowledge graph update algorithm

[0014] By comparing the changes between the old and new data, only the changed parts of the knowledge graph are updated. The graph difference algorithm is used to identify the nodes and edges that need to be updated in the knowledge graph. The difference in graph or node representation is calculated by embedding vectors. For the node embeddings of two graphs, the difference between the nodes is calculated using the following Euclidean distance or cosine similarity:

[0015]

[0016] Among them, Δh(v) is the difference in embedding vectors of node v in the new and old versions of the knowledge graph, is the embedding vector (low-dimensional vector representation) of node v in the old version of the knowledge graph, is the embedding vector of node v in the new version of the knowledge graph, ||·||: Euclidean distance L2 norm, used to measure the geometric distance between two vectors in space;

[0017] Finally, the dynamic update strategy based on reinforcement learning

[0018] Reinforcement learning is introduced to optimize the update strategy of the knowledge graph. By defining the reward function, the model can automatically learn when and how to update the knowledge graph. The deep Q network DQN method is used to train the update strategy. In Q learning, the goal is to learn a Q function. The update formula of Q learning is:

[0019]

[0020] Among them, Q(s t , a t ) is in state st Next, perform action a t The expected cumulative reward (Q value); α is the learning rate, 0≤α≤1, which controls the speed at which new information covers old values, r t To perform action a t The immediate reward obtained after γ is the discount factor 0≤γ≤1, balancing the importance of current rewards and future rewards. For the next state s t+1 The maximum expected Q value of all possible actions; s t+1 To perform action a t The next state to be transferred to;

[0021] (2) Multimodal information fusion

[0022] By combining multimodal data with knowledge graphs, multimodal encoding and graph neural networks are used to align semantic features and build a unified knowledge representation.

[0023] Multimodal feature extraction

[0024] Use pre-trained language models to extract semantic features of text and map them to entities and relations in the knowledge graph; use convolutional neural networks (CNNs) or visual transformers (ViTs) to extract visual features of images and associate them with entities in the knowledge graph; for video features, use 3D convolutional neural networks (3D convolutional neural networks) or video transformers to extract spatiotemporal features of videos and associate them with events and scenes in the knowledge graph;

[0025] Multimodal feature fusion

[0026] By calculating the attention weights between different modalities, the fusion ratio of each modal feature is dynamically adjusted; multimodal features are integrated into the knowledge graph using graph neural networks to generate a unified semantic representation;

[0027] Multimodal knowledge graph construction

[0028] Construct a multimodal knowledge graph, uniformly represent multi-source heterogeneous information as nodes and edges in the graph, forming a structured semantic network; through graph embedding technology, map the entities and relationships in the graph into a low-dimensional vector space, thereby achieving deep fusion and efficient representation of multimodal information;

[0029] (3) Solution to cold start problem

[0030] Combining users' social relationships and item context information, we build a more robust and scalable recommendation system. Specifically, we:

[0031] (1) Semantic reasoning based on knowledge graph

[0032] Use semantic information in the knowledge graph to generate initial recommendations for new users and new items, and use graph reasoning algorithms to perform semantic reasoning in the knowledge graph to generate cold start recommendation results.

[0033] (2) Context-based cold start recommendation

[0034] A context-aware cold-start recommendation system is used. The system integrates multi-dimensional context features to build a personalized recommendation mechanism. Specifically, the system uses a deep context embedding method to transform discrete context features into continuous low-dimensional vector representations, which are then deeply integrated with the structured semantic information in the knowledge graph.

[0035] (3) Cold start solution based on transfer learning

[0036] Use transfer learning to transfer the knowledge of existing users and items to new users and new items to alleviate the cold start problem; design a cross-domain transfer learning model to achieve knowledge transfer between different domains by sharing semantic information in the knowledge graph.

[0037] An intelligent recommendation algorithm based on a knowledge graph uses the system described above. This algorithm not only considers the user's historical behavior patterns, but also incorporates the user's interest evolution trajectory and multimodal data. Specifically, it includes:

[0038] In terms of graph neural network recommendation algorithms, we design an innovative recommendation framework that uses graph convolutional networks (GCNs) or graph attention networks (GAs) to propagate deep information in knowledge graphs. We use graph embedding technology to map users and items into a unified low-dimensional vector space and generate personalized recommendations by accurately calculating the similarity between them, thereby achieving more accurate recommendation services.

[0039] Reinforcement Learning (RL) is introduced to optimize the recommendation strategy. By defining a reward function, the model can automatically learn how to generate the best recommendation results. Deep RL method is used to combine the semantic information in the knowledge graph to generate personalized recommendation strategies.

[0040] The present invention has the following beneficial effects:

[0041] The real-time and accurate nature of recommendations has been significantly improved. The knowledge graph's dynamic update mechanism ensures the recommendation system can keep pace with user interests and external environmental changes, providing users with recommendations that better meet their current needs. Compared to traditional recommendation systems, this significantly improves recommendation accuracy, reducing user information sifting time and enhancing the user experience in shopping and browsing on the platform.

[0042] Enhanced personalized recommendation capabilities: Multimodal information fusion technology comprehensively captures user interests, resulting in recommendations that are more tailored to individual needs. In e-commerce scenarios, the system can recommend products that best meet users' aesthetic and functional needs. In content scenarios, it can precisely push articles and videos that match users' interests and preferences, improving user satisfaction and platform engagement.

[0043] Effectively alleviates the cold start problem: The cold start solution based on knowledge graph semantic reasoning, context perception and transfer learning provides reasonable initial recommendations for new users and new items, helps new users quickly discover content of interest, helps new items open up the market, expands the coverage of the recommendation system, and improves the overall recommendation performance of the system.

[0044] Optimized recommendation algorithm performance: This optimization, based on graph neural networks and reinforcement learning, combines multiple information sources to generate more accurate recommendations. This algorithm has significantly improved key metrics such as accuracy and recall, enabling it to better adapt to diverse user needs and dynamically changing business scenarios, thereby promoting the widespread application and in-depth development of intelligent recommendation systems across various fields. DETAILED DESCRIPTION

[0045] The specific implementation of the present invention will be further described below with reference to specific embodiments:

[0046] Intelligent recommendation system based on knowledge graph, including

[0047] (1) Dynamic update mechanism of knowledge graph. In order to enable the recommendation system to respond quickly to the latest needs of users and improve the real-time and accuracy of recommendations, it is necessary to design a knowledge graph dynamic update algorithm. It can track user behavior in real time and keenly capture changes in the external environment to ensure that the recommendation system always meets the actual needs of users. Specifically, it includes:

[0048] First, build an event-driven dynamic update framework

[0049] Design an event-driven knowledge graph update framework that triggers knowledge graph updates by monitoring user behavior (such as clicks, purchases, and ratings) and external environment changes (such as hot topics and seasonal trends). Use stream processing technologies (such as Apache Kafka and Flink) to capture user behavior data in real time and use it as input for knowledge graph updates.

[0050] Secondly, the incremental knowledge graph update algorithm

[0051] We propose an incremental update algorithm that compares the changes between old and new data and updates only the changed parts of the knowledge graph, avoiding the computational overhead of a full update. We use a graph differential algorithm to identify the nodes and edges that need to be updated in the knowledge graph, ensuring the efficiency and accuracy of the update. The difference between the graph or node representation can be calculated using embedding vectors. For the node embeddings of two graphs, the difference between the nodes can be calculated using the following Euclidean distance or cosine similarity:

[0052]

[0053] Among them, Δh(v) is the difference in embedding vectors of node v in the new and old versions of the knowledge graph, is the embedding vector (low-dimensional vector representation) of node v in the old version of the knowledge graph, is the embedding vector of node v in the new version of the knowledge graph, ||·||: Euclidean distance L2 norm, used to measure the geometric distance between two vectors in space;

[0054] Finally, the dynamic update strategy based on reinforcement learning

[0055] Reinforcement learning (RL) is introduced to optimize the knowledge graph update strategy. By defining a reward function (such as user click-through rate, purchase conversion rate, etc.), the model can automatically learn when and how to update the knowledge graph. The deep Q-network (DQN) method is used to train the update strategy to ensure that the knowledge graph update can maximize the performance of the recommendation system. In Q-learning, the goal is to learn a Q function. The update formula of Q-learning is:

[0056]

[0057] Among them, Q(s t , a t ) is in state s t Next, perform action a t The expected cumulative reward (Q value); α is the learning rate, 0≤α≤1, which controls the speed at which new information covers old values, r t To perform action a t The immediate reward obtained after γ is the discount factor 0≤γ≤1, balancing the importance of current rewards and future rewards. For the next state s t+1 The maximum expected Q value of all possible actions; s t+1 To perform action a t The next state to be transferred to;

[0058] (2) Multimodal information fusion

[0059] A multimodal information fusion method is proposed. By combining multimodal data such as text, images, and videos with knowledge graphs, multimodal encoding and graph neural networks are used to align semantic features to build a unified knowledge representation, thereby improving the personalization ability of the recommendation system.

[0060] Multimodal feature extraction

[0061] Use pre-trained language models (such as BERT, GPT) to extract semantic features of text and map them to entities and relations in the knowledge graph; use convolutional neural networks (CNN) or visual transformers (ViT) to extract visual features of images and associate them with entities in the knowledge graph; for video features, use 3D convolutional neural networks (3DCNN) or video transformers to extract spatiotemporal features of videos and associate them with events and scenes in the knowledge graph.

[0062] Multimodal feature fusion

[0063] A cross-modal attention mechanism is proposed. By calculating the attention weights between different modalities, such as text, images, and videos, it dynamically adjusts the fusion ratio of each modal feature. A graph neural network (GNN) is used to fuse multimodal features into the knowledge graph to generate a unified semantic representation.

[0064] Multimodal knowledge graph construction

[0065] This paper proposes building a multimodal knowledge graph that uniformly represents heterogeneous information from multiple sources, such as text, images, and videos, as nodes and edges within the graph, forming a structured semantic network. Using graph embedding technology, the entities and relationships within the graph are mapped into a low-dimensional vector space, enabling deep fusion and efficient representation of multimodal information, providing high-quality semantic features for subsequent recommendation computations.

[0066] (3) Solution to cold start problem

[0067] Combining users' social relationships and item context information, we build a more robust and scalable recommendation system. Specifically, we:

[0068] (1) Semantic reasoning based on knowledge graph

[0069] We use semantic information in the knowledge graph to reason and generate initial recommendations for new users and items. For example, for new users, we can infer their interests through their social relationships (such as friends and people they follow); for new items, we can infer their potential audience through information such as their category and brand. We use graph reasoning algorithms (such as random walks and graph attention networks) to perform semantic reasoning in the knowledge graph and generate cold-start recommendation results.

[0070] (2) Context-based cold start recommendation

[0071] The study proposes a context-aware cold-start recommendation system that integrates multi-dimensional contextual features (including spatiotemporal information, device status, etc.) to build a personalized recommendation mechanism. Specifically, the system uses a deep context embedding method to transform discrete contextual features into continuous low-dimensional vector representations, which are then deeply integrated with the structured semantic information in the knowledge graph to achieve more accurate recommendations.

[0072] (3) Cold start solution based on transfer learning

[0073] Transfer learning is used to transfer knowledge from existing users and items to new users and items, alleviating the cold start problem. A cross-domain transfer learning model is designed to enable knowledge transfer between different domains by sharing semantic information in the knowledge graph.

[0074] An intelligent recommendation algorithm based on a knowledge graph uses the system described above. This algorithm not only considers the user's historical behavior patterns, but also incorporates the user's interest evolution trajectory and multimodal data. Specifically, it includes:

[0075] In terms of graph neural network recommendation algorithms, we design an innovative recommendation framework that uses graph convolutional networks (GCNs) or graph attention networks (GAs) to propagate deep information in knowledge graphs. We use graph embedding technology to map users and items into a unified low-dimensional vector space and generate personalized recommendations by accurately calculating the similarity between them, thereby achieving more accurate recommendation services.

[0076] Reinforcement Learning (RL) is introduced to optimize the recommendation strategy. By defining a reward function (such as user click-through rate, purchase conversion rate, etc.), the model can automatically learn how to generate the best recommendation results. Using the Deep RL method, combined with the semantic information in the knowledge graph, a personalized recommendation strategy is generated.

[0077] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. Intelligent recommendation system based on knowledge graph, characterized by: include (1) Dynamic update mechanism of knowledge graph, including: First, build an event-driven dynamic update framework Design an event-driven knowledge graph update framework that triggers knowledge graph updates by monitoring user behavior and external environment changes. Use stream processing to capture user behavior data in real time and use it as input for knowledge graph updates. Secondly, the incremental knowledge graph update algorithm By comparing the changes between the old and new data, only the changed parts of the knowledge graph are updated. The graph difference algorithm is used to identify the nodes and edges that need to be updated in the knowledge graph. The difference in graph or node representation is calculated by embedding vectors. For the node embeddings of two graphs, the difference between the nodes is calculated using the following Euclidean distance or cosine similarity: Among them, Δh(v) is the difference in embedding vectors of node v in the new and old versions of the knowledge graph, is the embedding vector (low-dimensional vector representation) of node v in the old version of the knowledge graph, is the embedding vector of node v in the new version of the knowledge graph, ||·||: Euclidean distance L2 norm, used to measure the geometric distance between two vectors in space; Finally, the dynamic update strategy based on reinforcement learning Reinforcement learning is introduced to optimize the update strategy of the knowledge graph. By defining the reward function, the model can automatically learn when and how to update the knowledge graph. The deep Q network DQN method is used to train the update strategy. In Q learning, the goal is to learn a Q function. The update formula of Q learning is: Among them, Q(s t , a t ) is in state s t Next, perform action a t The expected cumulative reward (Q value); α is the learning rate, 0≤α≤1, which controls the speed at which new information covers old values, r t To perform action a t The immediate reward obtained after γ is the discount factor 0≤γ≤1, balancing the importance of current rewards and future rewards. For the next state s t+1 The maximum expected Q value of all possible actions; s t+1 To perform action a t The next state to be transferred to; (2) Multimodal information fusion By combining multimodal data with knowledge graphs, multimodal encoding and graph neural networks are used to align semantic features and build a unified knowledge representation. Multimodal feature extraction Use pre-trained language models to extract semantic features of text and map them to entities and relations in the knowledge graph; use convolutional neural networks (CNNs) or visual transformers (ViTs) to extract visual features of images and associate them with entities in the knowledge graph; for video features, use 3D convolutional neural networks (3D convolutional neural networks) or video transformers to extract spatiotemporal features of videos and associate them with events and scenes in the knowledge graph; Multimodal feature fusion By calculating the attention weights between different modalities, the fusion ratio of each modal feature is dynamically adjusted; multimodal features are integrated into the knowledge graph using graph neural networks to generate a unified semantic representation; Multimodal knowledge graph construction Construct a multimodal knowledge graph, uniformly represent multi-source heterogeneous information as nodes and edges in the graph, forming a structured semantic network; through graph embedding technology, map the entities and relationships in the graph into a low-dimensional vector space, thereby achieving deep fusion and efficient representation of multimodal information; (3) Solution to cold start problem Combining users' social relationships and item context information, we build a more robust and scalable recommendation system. Specifically, we: (1) Semantic reasoning based on knowledge graph Use semantic information in the knowledge graph to generate initial recommendations for new users and new items, and use graph reasoning algorithms to perform semantic reasoning in the knowledge graph to generate cold start recommendation results. (2) Context-based cold start recommendation A context-aware cold-start recommendation system is used. The system integrates multi-dimensional context features to build a personalized recommendation mechanism. Specifically, the system uses a deep context embedding method to transform discrete context features into continuous low-dimensional vector representations, which are then deeply integrated with the structured semantic information in the knowledge graph. (3) Cold start solution based on transfer learning Use transfer learning to transfer the knowledge of existing users and items to new users and new items to alleviate the cold start problem; design a cross-domain transfer learning model to achieve knowledge transfer between different domains by sharing semantic information in the knowledge graph.

2. An intelligent recommendation algorithm based on knowledge graph, using the system according to claim 1, characterized in that: The algorithm not only considers the user's historical behavior patterns, but also incorporates the user's interest evolution trajectory and multimodal data, including: In terms of graph neural network recommendation algorithms, we design an innovative recommendation framework that uses graph convolutional networks (GCNs) or graph attention networks (GAs) to propagate deep information in knowledge graphs. We use graph embedding technology to map users and items into a unified low-dimensional vector space and generate personalized recommendations by accurately calculating the similarity between them, thereby achieving more accurate recommendation services. Reinforcement Learning (RL) is introduced to optimize the recommendation strategy. By defining a reward function, the model can automatically learn how to generate the best recommendation results. Deep RL method is used to combine the semantic information in the knowledge graph to generate personalized recommendation strategies.

Images