A Novel Multi-modal Recommender with Modality-Specific Graph Refining Strategy

Recommender system (RecSys) have become essential in various domains, leveraging user-item interaction data to provide personalized suggestions. Collaborative filtering (CF) techniques, while effective, are often limited by the sparsity of historical data. To address this issue, multi-modal recommender systems (MMRec) integrate visual, audio, and textual information to enrich user and item representations, thus improving recommendation accuracy. However, recent advancements in GCNs-based MMRec merely constructed identical bipartite graphs for each modality, ignoring the semantic gap among different modalities, and eventually leading to sub-optimal recommendation.In this work, we propose a novel Multi-Modal Light Graph-Refining Convolutional Network (MMLGRN) recommender that constructs distinct user-item graphs for each modality. Specifically, MMLGRN employs pruning strategies to refine graph structures based on modality features. By learning modality-specific collaborative signals from different graphs and aggregating multi-modal embeddings, MMLGRN is able to deliver customized recommendations that cater to user’s preferences. Experimental results on a public dataset demonstrate that MMLGRN outperforms state-of-the-art baselines, validating the effectiveness of the proposed graph refining strategy.