Feature-level attention network with group-aware interest modeling for sequential recommendation

Abstract

Sequential recommendation focuses on modeling user preferences based on their historical interaction sequences to predict future behaviors with greater precision. Incorporating feature-level information beyond item IDs has become a crucial approach to improving the performance of the recommendation system. However, existing methods overlook the hierarchical group relationships among users. This limitation prevents these methods from fully capturing user preferences, leading to an incomplete understanding of their true interests. Meanwhile, effectively leveraging multi-source information in recommendation systems remains a significant challenge. Existing methods typically rely on simple techniques such as pooling or concatenation to integrate information from different sources, which could degrade overall performance. To address these limitations, we propose a novel approach: Feature-level Attention Network with Group-aware Interest Modeling for Sequential Recommendation (FANGIM). Specifically, we first employ two distinct encoders to generate user embeddings at different level. Next, we introduce a group clustering module, which identifies potential interest groups at multiple granularities and derives user group interest embeddings for both item and feature level interactions. Furthermore, we design a multi-source representation fusion module that effectively integrates information from diverse sources, reducing the semantic gap between different representation spaces. Additionally, we incorporate contrastive learning within this module to ensure consistency between the different levels of representations. Finally, extensive experiments demonstrate that FANGIM outperforms state-of-the-art baselines across four datasets.