Learning Potential Cross-Modal Interactions with Adaptive Relation Refinement for Multimodal Recommendation

Recent advancements in multimodal recommendation systems have demonstrated promising performance by fusing explicit interaction behaviors and multimodal features to model user preferences. Prevailing methods predominantly rely on explicit interactions yet fundamentally overlook the essential exploration of semantic-aware potential interactions. This limitation restricts existing methods to information-constrained interaction graphs. The introduction of an information-rich potential interaction paradigm can effectively extend user connections to more relevant items beyond interaction-based connections, thereby enhancing recommendation performance. To this end, we develop a novel framework called L earning P o tential Cro ss- M odal I nteractions (LoCroMI), which integrates collaborative and cross-modal signals to construct three distinct graphs: a potential user-item interaction graph and two interest-aware item graphs. We utilize graph convolutional networks for aggregation and introduce a dual optimization mechanism that combines contrastive learning with cross-modal interaction constraints to enhance the discriminability of representations. Furthermore, we devise an adaptive potential interaction refinement strategy, which dynamically balances noise suppression and hard potential interaction discovery via a confidence-based threshold during iterative training. LoCroMI establishes a robust framework that significantly improves the modeling of semantically rich user-item relationships. Extensive experiments on five benchmark datasets demonstrate that LoCroMI outperforms existing state-of-the-art methods. Code is available at: https://github.com/ZhuoxinHe/LoCroMI.