MIMOSA: A unified multi-view multi-order contrastive learning framework for bundle recommendation

Abstract

Bundle recommendation has emerged as a vital online service that provides users with personalized collections of items likely to attract them. While existing methods attempt to integrate multi-view information, they often struggle to effectively leverage associations between entities (i.e., users and bundles), particularly in distinguishing between associated and non-associated entities. To address this, our study proposes an innovative bundle recommendation approach termed multi-view multi-order contrastive learning (MIMOSA), which simultaneously models the underlying relationships across multiple views and between different entities. The approach introduces a unified graph-based contrastive learning framework that organizes both intra- and inter-entity associations using different orders of proximity within the user-bundle graph. Specifically, MIMOSA customizes tactics for positive sample discovery and contrastive loss calculation to capture the heterogeneous semantics of various entity associations. This enables the alignment of associated entity embeddings while effectively dispersing non-associated ones. Additionally, a center-matching strategy is designed to efficiently coordinate multi-view entity representations, thereby accelerating the contrastive learning process. Extensive experiments on two large-scale datasets, Youshu and NetEase, demonstrate MIMOSA’s superior performance over baseline methods. The results show that compared to the best baseline, our proposed approach achieves average improvements of 2.87% (R@20) and 3.02% (N@20) on Youshu, and 5.19% (R@20) and 4.07% (N@20) on NetEase.