Disentangled Sparse Graph Attention Networks with Multi-Intent Fusion for Session-based Recommendation

Session-based Recommendation has garnered considerable interests recently due to providing personalized recommendations based on anonymous behavior sequences. Most of existing models learn user preferences from a holistic perspective without delving into the key factors that drive user–item interactions. This results in an insufficient capture of user intent, as they rely solely on session data for predictions. Additionally, these models are more susceptible to the negative impact of noise on account of the limited short-term interactions. To tackle these problems, we propose a novel model called Disentangled Sparse Graph Attention Networks with Multi-Intent Fusion for Session-based Recommendation to learn item embeddings from factor level and model user intent for better inferring the user preferences. Specifically, we map item embeddings into multiple factors using disentanglement techniques and utilize gated graph neural network to learn the embeddings based on the item adjacent similarity matrix calculated for each factor. An innovative position information generation module is designed to encode the order of items. Subsequently, we model user intent from three perspectives and apply intent-aware fusion module to integrate them into a unified intent representation for incorporating intent information into the current session. Sparse attention networks are employed to denoise and extract the intent pattern of the current session. Furthermore, sessions exhibiting similar intent pattern are identified to augment the representation of the current session. Extensive experiments on five datasets indicate that our model outperforms state-of-the-art methods consistently.