IAMRec: Intent-Adaptive Multimodal Recommendation with Collaborative–Modality Disentanglement

Multimodal recommender systems have become increasingly important to leverage rich semantic signals from textual and visual content for effective recommendation. Despite significant progress, most existing approaches still encounter two major challenges: (C1) Feature homogenization from collaborative-modality coupling, when collaborative filtering signals and modality-specific features are entangled within shared user embedding learner; and (C2) Inadequate modeling of historical–candidate semantic relevance, when existing uniform aggregation fails to capture dynamic user preference intention variations. To address these, we propose an with collaborative-modality disentanglement, named . Specifically, the proposed IAMRec contains a (1) collaborative-modality representation disentangler to separately capture structural collaborative signals and semantic modality features; and an (2) intent-adaptive preference scorer to build item-sensitive user preferences via relevant historical relevance modeling, where a collaborative-modality alignment strategy is also introduced to ensure cross-modal consistency. Extensive experiments on three public multimodal datasets could verify the consistently superior performance of our proposed method over existing baseline models. The code is available at https://github.com/jiayichen1121/IAMRec .