Image and text recommendation system based on multimodal feature fusion

With the growing importance of multimodal data in recommendation systems, effectively integrating heterogeneous information—such as user behavior, text, and images—remains a key challenge. This paper proposes GFSAM (Graph-Fusion with Samformer for Multimodal Recommendation), a framework that enhances semantic integrity and training stability through structural modeling and optimization. GFSAM incorporates a structure-aware graph fusion module for aligning multimodal features via role embedding and adjacency perception, and integrates the Samformer module—combining RevIN and channel attention—to improve cross-modal representation and expression depth. Experiments on three datasets show that GFSAM achieves average gains of 3.86% in Recall and 5.42% in NDCG, demonstrating superior adaptability, robustness in sparse or heterogeneous scenarios, and improved training efficiency (e.g., 2.89 seconds per iteration on the Sports subset).