Pareto selective error feedback suppression for popularity-diversity balanced session-based recommendation

Yanglei Gan, Qiao Liu, Da Luo, Rui Hou, Yuxiang Cai, Run Lin, Xiaojun Shi

Abstract: Session-based recommendation approaches have garnered substantial attention for their ability to deliver tailored recommendations across diverse domains. However, balancing recommendation accuracy and diversity remains a significant challenge, with most research focusing primarily on accuracy at the expense of diversity. In this paper, we address this challenge by introducing a novel approach, the Pareto selective error feedback suppression algorithm (Pesa). This model-agnostic back-propagation algorithm simultaneously enhances recommendation accuracy and diversity by strategically modifying gradient distributions during training. Specifically, Pesa operates by randomly freezing the gradients of diverse items during back-propagation, thereby suppressing error feedback selectively and enabling the model to better capture critical features of these items while maintaining a balance between accuracy and diversity. Comprehensive evaluations on four benchmark datasets demonstrate the superiority of Pesa, obtaining relative improvement in accuracy, while achieving a 10% increase in diversity compared to state-of-the-art baselines. These results highlight the effectiveness of our approach in addressing the trade-off between accuracy and diversity in session-based recommendation systems. We provide our source code at: https://github.com/Jackymeister/Pesa.