Collaborative association networks with cross-level attention for session-based recommendation

Session-based recommendation aims to predict the next interacted item based on the anonymous user’s behavior sequence. The main challenge lies in how to perceive user preference within limited interactions. Recent advances demonstrate the advantage of utilizing intent represented by combining consecutive items in understanding complex user behavior. However, these methods concentrate on the diverse expression of intents enriched by considering consecutive items with different lengths, ignoring the exploration of complex transitions between intents. This limitation makes intent transfer unclear in the user behavior with dynamic change, resulting in sub-optimal performance. To solve this problem, we propose novel collaborative association networks with cross-level attention for session-based recommendation (denoted as CAN4Rec), which simultaneously models intra- and inter-level transitions within hierarchical user intents. Specifically, we first construct two levels of intent, including individual-level and aggregated-level intent, and each level of intent is obtained based on sequential transitions. Then, the cross-level attention mechanism is designed to extract inter-transitions between different levels of intent. The captured inter-transitions are bi-directional, containing from individual-level to aggregated-level intents and from aggregated-level to individual-level intents. Finally, we generate directional session representations and combine them to realize the prediction of the next item. Experimental results on three public benchmark datasets demonstrate that the proposed model achieves state-of-the-art performance.