Unifying multi-associations through hypergraph for bundle recommendation

Bundle recommendation, which seeks to recommend a group of items to users, is widely used in real-world applications. Despite the success of current bundle recommendation approaches, there are still significant challenges: (1) Multiple associations (e.g., user-bundle interactions, bundle-item affiliations, etc.) in bundle recommendation. (2) People's interactions with a single item and a bundle follow different patterns, e.g., users can quickly decide whether to purchase an item, but hard to determine with a bundle regardless of what items are included. (3) The data sparsity of user-bundle historical interactions. This paper proposes a Unified Hypergraph framework for Bundle Recommendation (UHBR) to tackle the aforementioned challenges. Specifically, UHBR first unifies multiple associations among users, bundles, and items into hypergraph, a more flexible and scalable data graph structure. Second, this hypergraph architecture allows both direct and indirect user-bundle relationships through items to be efficiently and comprehensively represented as hyperedges. Third, we leverage the potential of indirect association to diminish the impact of user-bundle interaction scarcity. Experimental results on two real-world datasets show that UHBR outperforms the state-of-the-art baselines by 15.9% on Recall and 19.8% on NDCG. Experiments further indicate that UHBR can alleviate data sparsity dilemma and has the highest efficiency.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Bundle recommendation is widely used in real-world applications, but it faces challenges such as multiple associations, different interaction patterns between users and items/bundles, and data sparsity. To address these challenges, this paper proposes a Unified Hypergraph framework for Bundle Recommendation (UHBR), which can comprehensively represent the relationships among users, bundles, and items in a more flexible way.