Data augmentation to enhance session-based recommendation systems with incomplete implicit feedback
Yucheng Shang, Yusuke Goto · SICE Journal of Control Measurement and System Integration · 2025
Personalized recommendations can alleviate the information overload problem by personalizing recommendations based on user preferences. Recommendation models harness the interactive information between users and items to make personalized recommendations, and most existing methods model users based on their historical clicks or purchase records. However, real-world datasets, especially those based on implicit feedback, frequently suffer from noise and missing data, thereby limiting the effective modelling of user preferences and item representations. This paper proposes an innovative data augmentation technique to identify potential data omissions, thereby enhancing the overall integrity of datasets. To evaluate its effectiveness, we applied the proposed method to two distinct datasets, i.e. the MIND news recommendation dataset and a retail dataset. This allowed us to investigate the generalizability of the proposed approach across different domains, one involving fast-moving content and the other focusing on purchasing behaviour. The experimental results demonstrate that integrating the proposed method with a standard recommendation model significantly improved the performance, particularly in terms of the hit ratio and normalized discounted cumulative gain.