Leveraging Heterogeneous Graph Structures for Enhanced User-Content Recommendation in Multi-Behavior Scenarios

Abstract With the advancement of personalized recommendation systems aiming to accurately capture user preferences for reasonable recommendations, most traditional personalized recommendation algorithms are based on collaborative filtering methods. In recent years, although machine learning methods have made significant progress over traditional approaches, current methods generally only consider single interactions between users and content or single relationships among content, making it difficult to effectively extract complex collaborative signals from diverse user-content interactions. To address this issue, we propose a personalized recommendation algorithm based on a graph attention mechanism and heterogeneous information networks (HA-Rec). First, by constructing a heterogeneous graph that models multiple user-content interactions and various content relationships, the proposed method maximizes information preservation. Additionally, by leveraging a graph attention mechanism on the content side, the model extracts effective features from heterogeneous graphs. It also incorporates interaction time, user information, and content information for comprehensive fusion, enabling fine-grained modeling of user-content relationships and node representation learning. Finally, experiments on public datasets demonstrate that the proposed model achieves superior performance on personalized recommendation tasks.