Neighborhood-Enhanced Multimodal Collaborative Filtering for Item Cold Start Recommendation

The lack of interaction data of new items in recommendation systems leads to the problem of cold-start item recommendations. Current methods usually approximate content features of the items to interaction embeddings and then use content features for prediction. However, these methods typically learn content features with the same item’s interaction embeddings, resulting in sparse supervised signals for content features and suboptimal recommendation performance. To address this issue, we propose a novel approach called Neighborhood-Enhanced Multimodal Collaborative Filtering (NEMCF) that captures the neighborhood of content features and interaction embeddings in order to enrich the content features. Specifically, we first design a dual graph neural network to aggregate interaction embeddings and content features of co-occurring neighbors of items. Then, we use the aggregated interaction embeddings and content features of items to supplement the content features via contrastive learning. Finally, to enhance the robustness of NEMCF, we randomly employ item content features and interaction embeddings for recommendations. Extensive experimental studies on four benchmark datasets demonstrate that the NEMCF method outperforms several typical cold-start methods and achieves state-of-the-art results.