Aggregately Diversified Bundle Recommendation via Popularity Debiasing and Configuration-Aware Reranking

Hyunsik Jeon, Jongjin Kim, Jaeri Lee, Jong-eun Lee, U Kang

Abstract: How can we expose diverse items across all users while satisfying their needs in bundle recommendations? Diversified bundle recommendation is a crucial task since it leads to great benefits for both sellers and users. However, there have been no studies on aggregate diversity in bundle recommendation, while they have been intensively studied in item recommendation. Moreover, existing methods of aggregately diversified item recommendation are not fully suitable for bundle recommendation. In this paper, we propose PopCon (Popularity Debiasing and Configuration-aware Reranking), an accurate method for aggregately diversified bundle recommendation. PopCon mitigates the popularity bias of a recommendation model by a popularity-based negative sampling in training process, and maximizes accuracy and aggregate diversity by a configuration-aware reranking algorithm. We show that PopCon provides state-of-the-art performance on real-world datasets, achieving up to \(60.5\%\) higher Entropy@5 and \(3.92\times \) higher Coverage@5 with comparable accuracies compared to the best competitor.