Learning multi-behavior user intent for session-based recommendation

Session-based recommendation makes a recommendation by exploiting short-term user interaction and has become a hot research topic. A quantity of methods have been proposed for session-based recommendation. Although effective, these methods focus on a particular behavior or fail to model the complex relationship between items with multiple behaviors. Thus they can only provide limited recommendation performance. To solve the above problem, we design a Multi-behavior User Intent Recommendation model (MUIR) to produce a recommendation to explore the complicated item-item dependencies with multiple behavior to make more effective recommendation. Technically, MUIR captures item-item cross-session global dependencies and generates behavior-specific user intent representations through a behavior-aware global attention encoder. MUIR employs a low-rank self-attention network to model in-session local dependencies within a session. Moreover, a proposed user behavior discrimination task is auxiliary optimized to disentangle the user intent with multiple behaviors. The extensive experiments show that our MUIR consistently outperforms the various state-of-the-art baselines.