Unified Empirical Evaluation and Comparison of Session-based Recommendation Algorithms

Abstract

Recently, session-based recommendation systems (SBRSs) have become a highly explored area, and numerous methods have been proposed. The abundance of related work poses a challenge for newcomers in comprehending the current research landscape and burdens researchers during method validation. Offering a thorough research overview helps newcomers understand the current research. Additionally, comparing representative methods in a consistent environment allows researchers to streamline their workload by focusing on the top-performing methods. Existing theory-oriented review articles introduce the main techniques employed in SBRSs but lack a detailed exploration of their specific applications. The most recent neural method evaluated in existing experiment-driven review was published in 2019, and the latest state-of-the-art methods haven’t been included. To address these gaps, this paper offers a more thorough overview of SBRSs. Specifically, we first categorize and overview existing methods. Then, we introduce the main techniques and illustrate their applications. The performance of representative methods is validated under identical experimental conditions to ensure reliable comparative results. Our findings indicate that dataset characteristics significantly impact model performance, and attention mechanisms-based and gated neural networks (GNNs)-based models generally outperform others. Finally, we propose potential directions for future research in SBRSs.