Addressing Hybrid Confounder Issue for Causal Session-Based Recommendation

In the field of session-based recommendation, users’ interaction behaviors are affected by both popularity and surrounding environmental factors, such as item popularity, season, or salary, and so on. Utilizing causal learning techniques to mitigate the negative impact of hybrid confounders on user preferences is a pressing challenge in session-based recommendation. Aiming at the above problems, this paper proposes a kind of causal session-based recommendation model for addressing hybrid confounder issue (HCCSRec). First, the conditional probability about user preference given a user sequence is solved by intervention and structural equation. Then, implicit environmental features are learned with observation sequence data. Next, user preference for the next item is estimated from the observation sequence, environmental feature and item popularity. Finally, the objective function is constructed by cross entropy and KL scatter, and the model parameters are learned. We perform extensive experiments on three real-world datasets from Beauty, MovieLens, and Yelp. Empirical studies validate that compared with the state-of-art recommendation methods, the HCCSRec method is helpful to discover user real interests, and can further improve the recommendation accuracy.