A Graph Attention-Based Recommendation Framework for Sparse User-Item Interactions

Abstract

This paper addresses the common issue of sparse user-item interaction data in recommendation systems and proposes a recommendation algorithm based on the graph attention mechanism. The goal is to enhance modeling ability and recommendation performance in sparse scenarios. The method first constructs an interaction graph structure between users and items and utilizes the graph attention network for efficient information aggregation of user and item nodes. By adaptively assigning attention weights to neighboring nodes, the model uncovers higher-order semantic relationships and latent preference representations. In terms of model design, this paper introduces embedding update strategies and regularization mechanisms to control and optimize the inter-layer evolution of node representations and mitigate overfitting issues. The experiments are based on the Amazon Product Review dataset and include several comparative experiments, such as the impact of data sparsity, embedding update methods, and regularization techniques. These experiments comprehensively evaluate the performance of the proposed algorithm under different training setups. The results show that the proposed method outperforms the comparison models across multiple metrics, including F1 score and AUC, demonstrating stronger robustness and generalization ability. This work not only validates the effectiveness of the graph attention mechanism in sparse recommendation but also provides valuable insights into the integration of graph structure modeling and representation learning.