Modeling category and multi-level user intentions for session-based recommendation

Session-based recommendation is extensively used in online services and businesses, which predicts users’ next interest with anonymous behavior sequence. Past research on session-based recommendation capturing user preference commonly focused solely on extracting the diversity of user intentions and the identification information of interacted items. However, fine-grained user intentions, which are reflected by both individual and consecutive items in session sequences, and the side information, item categories, which can help to learn the potential user intentions in session sequences have not been explored deeply. These insights inspire us to model category and multi-level user intentions for session-based recommendation (CIS). Specifically, session sequences are first split pairwise between levels, and connected sequentially within levels to construct multi-level user intention graphs and item category graphs and model the fine-grained information. Moreover, we refine the representation of user intentions and item categories with graph convolutional network and iterative update process, in order to extract potential user intentions based on category semantics. Finally, we generate session embeddings based on each level of user intentions, and introduce an Integration Predicting strategy to anticipate users’ next interested item. Extensive experiment results on Diginetica dataset and Tmall dataset demonstrate that CIS is superior to other state-of-the-art baselines, with improvements of 24.38% in Hit Ratio (HR), 69.56% in Mean Reciprocal Rank (MRR) and 55.94% in Normalized Discounted Cumulative Gain (NDCG).