Multi-Modal Contrastive Fusion for Consensus Learning in Sequential Group Recommendation

Yue Kou, Dong Li, Qixiang Tang, Derong Shen, Tiezheng Nie

Abstract: A good group recommender system should be capable of modeling how a consensus among group members is reached. However, existing works fail to adequately capture the dynamic consensus that underlies group members’ behavioral sequences for effective sequential group recommendation. In this paper, we propose a multi-modal Contrastive Fusion framework to learn dynamic consensus for Sequential Group Recommendation (CF-SGR). Specifically, we utilize user-item interactions, group-item interactions, and side information to encode both collaborative and semantic signals, further learning the group consensus in these different modalities. Subsequently, we employ a multi-modal contrastive fusion strategy to effectively integrate the latent embeddings of these modalities. We also propose a strategy for generating explanations, which integrates LLMs with CPT (Continuous Prompt Tuning). Our method outperforms state-of-the-art group recommendation methods on two real-world datasets.

External IDs:dblp:conf/icmcs/KouLTSN25