Multi-level category-aware graph neural network for session-based recommendation

Abstract

Graph neural network (GNN)-based models have achieved state-of-the-art performance in session-based recommendation (SBR). In our research, we observe that the subsequence-level intra-category item–item transition can reflect user preference under specific category, thus enhancing recommendation performance. This is overlooked by some existing category-aware SBR models. Moreover, most existing SBR models focus solely on item prediction. However, we find that item prediction and category prediction are closely related, which can respectively serve as the main task and the auxiliary task within a joint multi-task learning (MTL) framework to enhance the performance of the main task. In this paper, we propose an SBR model named Multi-level Category-aware Graph Neural Network (MCGNN) that captures multi-level information involved in a session including subsequence-level intra-category item–item transition, and performs both item prediction and category prediction in an MTL manner. To achieve this, we propose to construct an intra-category graph for each session to model subsequence-level intra-category item–item transition, sequence-level category–category transition and item–category relation. We also construct a sequence graph for each session to model sequence-level item–item transition. These two graphs capture the multi-level information involved in a session. Next, we apply GNNs to learn item representations and category representations through the two graph views. Finally, we perform item prediction and category prediction in a joint MTL framework. Our experimental results on three widely used datasets show the superiority of our proposed MCGNN.