Sentiment-Time Heterogeneous Residual Graph Attention Transformer for Session-Based Recommendation

Abstract

Session-Based Recommendation (SBR) systems are facing considerable challenges, with their primary objective being to implement precise recommendations based on users’ historical behavior sequences. Graph Neural Networks (GNNs) have emerged as powerful tools for processing graph-structured data in recommendation systems. Although recent research has advanced in this area, a significant gap remains in the in-depth exploration of transitional relationships between user interests. Additionally, real-world recommendation scenarios typically involve various heterogeneous relationships, which contain a wealth of information that can significantly enhance the learning of user preferences. To address this research gap, in this paper, we introduce a model termed the Sentiment-Time Heterogeneous Residual Graph Attention Transformer (STH-ResGAT), which is designed to capture the dynamic nature of user interests and the complexity inherent in heterogeneous graphs. In STH-ResGAT, we develop a Sentiment-Time-Heterogeneous Graph (STH-Graph) that integrates sentiment and time factors into the edges of the graph structure. Furthermore, the Residual Graph Attention Transformer for Heterogeneous Networks (ResGAT-Het) is designed to manage diverse node and edge types based on the STH-Graph. Extensive experiments are conducted on four widely-used benchmark datasets: Ciao, Yelp, Epinions and LibraryThing. The results demonstrate that our proposed STH-ResGAT method significantly outperforms previous state-of-the-art baseline approaches.The implementation of ResGAT-Het is available in https://github.com/zhangyu2234/ResGAT-Het.git .