AMMRM: an adaptive multi-modal recommendation model with noise filtering and modal feature enhancement

The effective integration of multimodal information is the key to improving the performance of recommendation systems. However, the common noise interference in multimodal data, such as irrelevant visual details or redundant text descriptions, can seriously affect the accurate modeling of items. Meanwhile, the preference of recommendation systems for popular items often leads to homogenization of recommendation results. To address these challenges, this paper proposes an Adaptive Multimodal Recommendation Model (AMMRM) that integrates noise filtering and feature enhancement. This model innovatively designs a multimodal noise-filtering gate, effectively reducing noise interference; Simultaneously designing a degree-sensitive edge-pruning strategy significantly alleviates the overfitting problem caused by popular items. AMMRM enhances the multimodal representation of items through a cross-modal multi-head attention mechanism and optimizes modal vectors using behavior-guided networks. In addition, AMMRM utilizes Graph Convolutional Networks (GCNs) to construct modal level user-item interaction graphs, deeply mining users’ potential preferences, and achieving fine integration of user preferences through adaptive fusion gates. The experimental results show that AMMRM outperforms the existing state-of-the-art baseline models on three public datasets (Baby, Sports, Clothing). Under AMMRM, Recall@20 has increased by 2.52%, 3.88%, and 3.80%; NDCG@20 has increased by 8.43%, 5.04%, and 3.03%, respectively. Future research will explore the use of knowledge graphs to enrich item representations and further enhance the performance of recommendation systems.