DASGRec: Disentangled and Aligned Structure-aware Graph Learning for Multimodal Recommendation
Abstract
Multimodal recommender systems leverage heterogeneous content such as images and text to enhance user preference modeling. However, existing methods suffer from two fundamental limitations: (1) they inadequately model the compositional nature of multimodal semantics, often overemphasizing modality-shared information while under-exploring modality-specific and cross-modal synergistic signals; and (2) they remain limited in handling heterogeneous multimodal noise, which is entangled with useful information and further propagates through graph-based interaction modeling. To address these issues, we propose Disentangled and Aligned Structure-aware Graph Learning for Multimodal Recommendation (DASGRec), a unified framework that systematically integrates semantic disentanglement, structure-aware graph learning, and information-theoretic optimization. Specifically, DASGRec first decomposes multimodal representations into shared and modality-specific components, enabling structured modeling of heterogeneous semantics. It then constructs semantic view-specific interaction graphs via affinity estimation and reliability-aware filtering, allowing different semantic components to induce distinct connectivity patterns. To coordinate these structures, we introduce a dependency-based structural alignment mechanism that promotes shared consistency, shared-specific disentanglement, and specific-view non-redundancy. Furthermore, we propose a synergy-oriented modeling module to capture interaction signals that emerge only through cross-modal coupling. Finally, we incorporate an information bottleneck objective to learn compact yet sufficient representations, effectively suppressing redundant and noisy information. Extensive experiments on multiple real-world datasets demonstrate that DASGRec consistently outperforms state-of-the-art baselines.