Balancing Diversity in Session-Based Recommendation Between Relevance and Unexpectedness

Abstract

Recommender systems encounter the potential problem of filter bubble, neglecting the diversity of recommendations. These systems are inevitable to lower user experience because they cannot but provide tedious recommendations. Although several solutions have been introduced to increase diversity, it is still challenging to prevent accuracy loss with diversity enhancement. This study presents a new user-oriented algorithm for session-based recommendations that aims to improve diversity in consideration of two serendipity components—relevance and unexpectedness. Specifically, our approach first adopts serendipitous preference embedding into the recommender system based on session and graph neural networks. Next, we leverage a greedy algorithm of the maximum a posteriori (MAP) inference for the determinantal point process to re-rank items. Lastly, it additionally incorporates personalized trade-off balancing through a parameter that can be controlled by the user. To validate our approach, we conducted an experiment with two real-world datasets to demonstrate its ability to balance accuracy and diversity. The results showed that our approach generated not only relevant but unexpected recommendations, successfully improving diversity without accuracy loss. This study contributes to recommendation diversification methods, especially for session-based recommender systems under the user-centric perspective.