Integration of short and long-term interests: A preference aware session-based recommender

Abstract

Session-based recommenders have gathered tremendous e-commerce and media streaming applications where the task is to predict the next item user would consume based on the session history. In this arena, advancements have been accomplished using deep learning techniques to model users’ long-term and short-term preferences in a session. The long-term module focuses on the entire item set in a session, while the short-term emphasizes a fixed set of items just before making a choice. While users’ short-term preferences in general reflect their immediate actions or intentions, consideration of such preferences, specifically with a fixed item set, makes the prediction task increasingly challenging. In this work, we attempt to capture the short-term behavior of users considering a dynamic window which self-tunes depending on the category of the last few items. A novel gating network that leverages the different drivers such as price, rating, dwell time, etc. and the latent representation of the items and their categories is introduced. The efficacy of our proposed method Gated Session-based Recommender integrating Short and Long-term preferences (GSRSL) has been evaluated against eight recent baselines on four publicly available real-world datasets, Tmall, Diginetica, and Cosmetics. Our findings demonstrate a significant improvement in MRR, Hit, and NDCG of up to 13.62%, 11.46%, and 13.00%, respectively, across all the datasets.