Session-Based Recommendation Method Using Popularity-Stratified Preference Modeling

Abstract

Large-scale offline evaluations of user–project interactions in recommendation systems are often biased due to inherent feedback loops. To address this, many studies have employed propensity scoring. In this work, we extend these methods to session-based recommendation tasks by refining propensity scoring calculations to reflect dataset-specific characteristics. We evaluate our approach using neural models, specifically GRU4REC, and K-Nearest Neighbors (KNN)-based models on music and e-commerce datasets. GRU4REC is selected for its proven sequential model and computational efficiency, serving as a robust baseline against which we compare traditional methods. Our analysis of trend distributions reveals significant variations across datasets, and based on these insights, we propose a hierarchical approach that enhances model performance. Experimental results demonstrate substantial improvements over baseline models, providing a clear pathway for mitigating biases in session-based recommendation systems.