Preference-Aware Multimodal Contrastive Learning Recommendation

Multimodal recommendation systems struggle to effectively integrate heterogeneous modalities and select informative negative samples, which limits their ability to capture users’ true preferences. To address these challenges, we propose Preference-Aware Multimodal Contrastive Learning (PAMCL), a novel framework that leverages a lightweight large language model (LLM) to generate explicit user preference descriptions for guiding the recommendation process. Our approach introduces a simplified multimodal graph aggregation mechanism that unifies feature propagation and contrastive learning, enhancing semantic consistency while reducing computational complexity. Additionally, we develop a preference-guided multiple negative sampling strategy that utilizes LLM-generated preference embeddings to identify challenging negative samples while avoiding false negatives. Experimental results on three public multimodal recommendation benchmarks demonstrate that PAMCL significantly outperforms state-of-the-art methods by effectively addressing false negative issues in sampling and noise contamination in multimodal fusion.