Recommendation method and device based on diffusion fusion of multimodal semantics and behavioral relationships

Abstract

The present invention discloses a recommendation method and device based on the diffusion fusion of multimodal semantics and behavioral relationships. The method comprises: training a diffusion model using a semantic item adjacency matrix and a behavioral item adjacency matrix; inputting the semantic item adjacency matrix into the trained diffusion model to generate a diffusion-aware item adjacency matrix; obtaining an original total item representation, a diffused total item representation, and a collaborative user representation; updating the initial user and item embeddings; obtaining an updated collaborative user representation and an updated diffused total item representation based on the updated user and item embeddings; and obtaining item recommendation results based on the updated collaborative user representation and the updated diffused total item representation. The method can obtain more complete structural information that truly reflects the relationships between items, reduces the impact of inherent multimodal noise, and significantly improves the accuracy of personalized recommendations.

Images