Beyond pairwise user transfer: Multimodal group-level relational preference modeling for cross-domain recommendation

Cross-domain recommendation (CDR) offers an effective way to alleviate the cold-start problem, but most existing methods still model cross-domain preference transfer at the level of individual users. This point-wise paradigm relies heavily on overlapping users and struggles to capture the higher-order relational patterns underlying user interests across domains, resulting in limited performance when overlap is scarce. Meanwhile, representing items solely with ID embeddings ignores rich multimodal content, further constraining transferable preference modeling. In this work, we propose Multimodal Group-Level Relational Preference Modeling for Cross-Domain Recommendation, a framework that reformulates the transferable unit from direct user-to-user embedding mappings to user-to-key-user relational signatures, where overlapping users provide supervision and non-overlapping users serve as relational anchors. This design captures more stable and transferable cross-domain interest structures, yielding consistently improved performance. Concretely, we employ a multimodal large language model to extract item representations from textual and visual content, and further adopt a temporal attention pooling encoder to integrate users’ historical interactions with multimodal item semantics for constructing multimodal user representations. Based on these enriched representations, we develop a group-level relation learning module that models relational dependencies between each overlapping user and a domain-specific key user set selected from the non-overlapping ones, thereby enabling group-level preference transfer across domains. We further introduce an adversarial alignment strategy to reduce cross-domain representation discrepancies and enhance generalization. Extensive experiments on real-world datasets demonstrate that our method consistently outperforms state-of-the-art baselines.