Disentangling Shared and Specific Representations for Multimodal Cross-Domain Recommendation via Cosine-Based Regularization

Cross-domain recommendation (CDR) leverages knowledge from a source domain to improve recommendation accuracy in a target domain with scarce data. Recent approaches attempt to disentangle user preferences into shared and domain-specific representations. However, most existing methods rely solely on simple user-item interaction information, making effective disentanglement challenging. Methods that employ contrastive learning to separate representations optimize relative distances, which may fail to suppress semantic overlap between representations and lead to inappropriate knowledge transfer. We propose a novel CDR method that integrates multimodal features and introduces a cosine similarity-based regularization term to suppress correlations between shared and domain-specific representations, ensuring each representation captures semantically distinct information. Our approach features: (1) rich representation learning using diverse modalities, (2) direct orthogonality constraints between representations via cosine similarity, and (3) explicit directional correlation suppression in the representation space. Experiments on large-scale Amazon datasets demonstrate our method outperforms state-of-the-art CDR approaches by an average of 3.7% over the best baseline in each domain. The code is available at https://github.com/meruemon/MMCDR.