Research on Recommendation Algorithm Based on Knowledge Graph

In response to issues such as data explosion leading to data overload and a subsequent decrease in the effectiveness of information retrieval, this paper proposes a Knowledge Graph Attention Network Splicing Semantics (KGAT-SS) model based on attention mechanisms. The model combines graph attention networks with a dual-tower model framework, extending the breadth of recommendations through the fusion of entity and text semantics. It enhances the depth of recommendations based on graph attention networks and adds constraints on the weights of transfer nodes in the graph to facilitate more efficient learning of embedded representations in nodes. The model consists of three modules: text processing, graph representation, and prediction. The main contributions include utilizing the pre-trained natural language processing model BERT for vectorizing user and item review texts, GRU encoding for further hidden information exploration, TransR mapping of instances in the dataset to vectors, and knowledge representation through the knowledge graph. The graph representation module employs graph attention networks to differentiate weights between nodes, allowing nodes to assess the importance of received information based on neighboring node weights. A threshold is set during propagation to filter out low-relevance entities. In the prediction layer, multiple representations of entity nodes are concatenated with semantic vectors of user and item texts from the text processing module to obtain the final vector representation. The matching degree is calculated through attention scores. Experimental results indicate that the proposed algorithm outperforms baseline models, leading to an improvement in recommendation effectiveness and enhancing the recommendation performance of the system.