PRIDES: A Probabilistic Model for Recurrent User Interest Drift Identification in Session-Based Recommendation

Abstract

A Session-Based Recommendation (SBR) identifies correlations among session interactions to understand user preferences and generate appropriate recommendations. A key challenge in this context is the dynamic change in user preferences, particularly when preferences disappear and reappear within a session, a phenomenon referred to as Recurrent User Interest Drift (RUID). Effectively capturing RUID is significant for aligning recommendations with ongoing user preferences. Existing SBR approaches often misclassify user preferences that differ from other session interactions as noise (unintentional interactions), relying on dwell time (the amount of time a user spends viewing an item) or neighboring sessions, thereby overlooking their potential reappearance as RUID later in the session. To the best of our knowledge, this work is the first to address the challenge of identifying RUID in SBR. The proposed approach assigns probabilistic scores to each interaction by considering its similarity to the immediate previous interaction, its inclusion among popular items (items with a higher number of interactions), its similarity to previous interactions, and the dwell time. As user preference may reappear anytime during the session, and RUID identification requires analyzing subsequent interactions, a list-based approach is used to retain these interactions until the session ends, enabling effective RUID identification. The matrix factorization-based attentive session encoder incorporates both short-term (ongoing) preferences and long-term (historical) preferences to generate personalized recommendations. Experimental results on three benchmark datasets, Yoochoose, Last.fm, and Gowalla, show that our method outperforms 14 state-of-the-art baselines, achieving an improvement of 2.28% in recall@20 and 1.39% in Mean Reciprocal Rank (MRR@20) on Yoochoose, 3.58% in recall@20 and 2.70% in MRR@20 on Last.fm, and 5.35% in recall@20 and 4.17% in MRR@20 on Gowalla datasets. © 2013 IEEE.