RMGAT: Bipartite Interaction Graph-Based Relation Fusion Multi-Head Graph Attention Recommendation Model

Abstract

Recommendation systems are crucial for accurately matching user needs in an environment with massive information. Traditional recommendation models have many deficiencies when dealing with user-item relationships and it is difficult to fully explore the information contained in user behaviors. This paper proposes a relation fusion multi-head graph attention recommendation model based on bipartite interaction graphs (RMGAT), aiming to innovate the performance of recommendation systems. The RMGAT model is based on the theory of user behavior, clearly dividing the interaction between users and items into explicit and implicit relationships, and constructing a bipartite interaction graph. By modeling these two types of relationships respectively, the explicit preferences with high credibility and the widely existing implicit latent interests can be effectively distinguished, overcoming the drawback of the traditional model in the semantic fuzzy processing of user behaviors. Meanwhile, the relational fusion multi-head graph attention mechanism is innovatively designed. With the help of the multi-head structure, the multiple semantic features of the explicit and implicit relationships are captured in parallel. The attention coefficient is used to dynamically adjust the weights to achieve the efficient complementary fusion of explicit and implicit information, significantly enhancing the model's expression ability for complex relationships, and thereby improving the accuracy and diversity of the recommendation results. After experimental verification on multiple typical datasets, the RMGAT model performs outstandingly in the recommendation scenarios of different fields and has achieved significant improvements in multiple key indicators compared with mainstream models. Future research will focus on developing lightweight multi-head mechanisms, introducing dynamic relationship modeling, and constructing cross-modal heterogeneous fusion frameworks, etc., in order to further optimize model performance and expand its application potential in more fields.