Multi-modal Negative Sampling for Recommendation with User Interest

Multi-modal Recommender Systems (MRSs) integrate various types of data, such as images, text, and videos, to deliver more accurate and personalized recommendations to users. However, traditional Multi-modal recommendation methods typically rely on random negative sampling strategies, overlooking the latent semantic relationships between items. Furthermore, these methods often suffer from the loss of user interest information during the denoising process. To address these challenges, this paper proposes a novel approach called Multi-modal Negative Sampling for Recommendation with User Interest (MNS-UI). To correct negative sample selection, MNS-UI introduces a dynamic negative sampler that incorporates Multi-modal information and allow for more precise capture of item semantic relationships. Additionally, we propose a reliable graph construction technique that preserves trusted edges in the user-item interaction graph, thereby mitigating the loss of user interest during the denoising process. Furthermore, we design a loss function to align Multi-modal features with item IDs, which significantly enhances the system's recommendation performance. Experimental results demonstrate that MNS-UI significantly outperforms state-of-the-art methods across three real-world datasets, with an average improvement of 2.4% in Recall@20, 6.7% in NDCG@20, and a reduction of 87.3% in the number of training epochs required for convergence.