<p>In recent years, multimodal recommendation has attracted significant attention, as it enables the modeling of user preferences by leveraging both user–item interaction data and multimodal item information. Although contrastive learning-based multimodal recommendation models have achieved promising results, they still suffer from challenges such as modal noise contamination and the entanglement of modal noise with user behavioral features, which hinder the accurate identification of user-item preference signals. To address these issues, we propose a Behavior-Guided Noise Purification for Multimodal Graph Contrastive Learning Recommendation Model (BGP-MGCL). Specifically, BGP-MGCL utilizes behavioral interaction information to filter out preference-irrelevant noise from the original item modalities. It then constructs a graph convolutional learner to generate multimodal user and item node representations based on neighboring nodes, and finally employs InfoNCE-based contrastive learning to capture the importance of different modalities, thereby enabling more accurate modeling of user–item preferences. We validate the effectiveness of the proposed BGP-MGCL on three multimodal recommendation datasets. Experimental results demonstrate that BGP-MGCL achieves significantly better performance compared to existing baseline methods.</p>