STAR: A session-based time-aware recommender system

Abstract

Session-Based Recommenders (SBRs) are designed to predict users’ next actions based on their previous interactions within a session, without access to historical information about users. Modern SBRs leverage deep neural networks to capture users’ current interests and map them to a latent space, enabling prediction of their next preference. While state-of-the-art SBR models achieve satisfactory results, they often overlook the temporal details of events within sessions, focusing instead on the sequence of events. To address this limitation, we propose the STAR framework, which incorporates session temporal information to enhance the performance of SBRs. By incorporating time intervals between events within sessions, we construct more informative representations for both items and sessions. Our mechanism revises session representation by embedding time intervals without using discretization. Empirical results on the Yoochoose and Diginetica datasets demonstrate that our proposed method outperforms state-of-the-art baseline models in Recall and MRR criteria. Our approach highlights the potential of session temporal information in enhancing the performance of SBRs by capturing the momentary interests of anonymous users and their mindset shifts during sessions.