Session Target Pair: User Intent Perceiving Networks for Session-Based Recommendation

Abstract

Session-based recommendation (SBR) aims to predict the next-interacted item based on an anonymous user behavior sequence (session). The main challenge is how to decipher the user intent with limited interactions. Recent progress regards the combination of consecutive items in the session as intent. However, these methods, which merely depend on the session, ignore the fact that such limited interaction within the session may not entirely express user intent. Therefore, it constrains the expression of diverse user intent without considering the candidate items to be predicted, which can be regarded as target intent, leading to a sub-optimal inference of user behavior. To solve the problem, we propose a novel I ntent A lignment N etwork for session-based recommendation ( IAN ), which models intent from both session and target perspectives. Specifically, we propose that session-level intent is explicitly formed by weighted aggregation of successive items, whereas target-level intent is composed of interacted and undiscovered items that are compatible. Based on it, we devise an intent alignment mechanism to ensure consistency between these two types of intent and obtain mutual intent representation. Finally, a gated mechanism is used to fuse mutual intent and target intent to generate session representation for prediction. Experimental results on three real-world datasets exhibit that IAN achieves state-of-the-art performance.