Prototype learning based hierarchical decoupling for multimodal recommendation

doi:10.1016/j.eswa.2025.130763

Prototype learning based hierarchical decoupling for multimodal recommendation

2025-12-13

0

PRE

AI

DOI：10.1016/j.eswa.2025.130763

Original

Original request for help

Save

Abstract

En 中文

Multimodal recommendation systems, which utilize heterogeneous information such as text and images to enhance user interest modeling, have become core components of online services. A common paradigm involves extracting multimodal features via encoders and integrating them into collaborative filtering (CF) frameworks. However, this integration process suffers from multi-level coupling: (1) Cross-path Coupling (where user ID embeddings are forcibly shared between the collaborative and modal paths, introducing noise propagation and distorting preference learning); (2) Inter-Modal Redundant Coupling (where extensive redundant information across modalities obscures discriminative knowledge, resulting in the fragmentation of critical information specific to modality).

Prototype learning based hierarchical decoupling for multimodal recommendation

Abstract

Journal

Organization