Session-based Recommendation Framework via Counterfactual Inference and Attention Network

Session-based recommendation systems (SRS) can predict the next action of anonymous users based on their historical behavior sequence, while some features directly influence whether an interaction occurs, which leads to inaccurate recommendation results that can't reflect the user's real preferences. This causes confounding effects in SRS, which causes the recommendation system to misunderstand user preferences and recommend unsatisfied items to users. To address this problem, we propose a session-based recommendation framework via counterfactual inference and attention networks (SRS-CIAN). This framework introduces external attention mechanisms into session recommendation tasks and combines causal graphs for modeling while capturing information on items within the session. We use counterfactual inference to refactor counterfactual scenarios for handling the confounding effects. Through extensive empirical experiments on real-world datasets, we demonstrate that our approach surpasses several strong baselines for confounding effects.