Category-integrated Dual-Task Graph Neural Networks for session-based recommendation

Yuhan Ding, Zizhuo Zhang, Bang Wang

Abstract: Session-based recommendation (SBR) aims to predict the subsequent item a user may be interested in based on their behavior within a limited timeframe. Most existing approaches primarily harness item relations and overlook the significance of attribute information (e.g. category). Users’ interests in specific items could change frequently within a single session, yet may exhibit more stability at the category level. We argue that integrating category information into SBR models can help mitigating data sparsity challenges for promoting next-item prediction. In this paper, we propose a novel SBR methodology named Category-integrated Dual-Task Graph Neural Networks (CDT-GNN). It constructs a heterogeneous global graph encompassing all sessions and individual heterogeneous local graphs for each session to learn items’ and categories’ representations. A dual-task learning strategy is employed to involve next-category prediction which serves as an auxiliary task to bolster the major task of next-item prediction. Additionally, a user-selectable accessory feature is developed to enhance the utilization of the predicted category. Extensive experimental results on three real-world datasets validate the effectiveness of CDT-GNN.