.png){kind=icon}
Return
A multi-level contrastive learning framework with reliability estimation for multimodal recommendation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.116352.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Existing multimodal recommendation systems often face content-level and interaction-level uncertainty: multimodal item content may contain contextual or weakly preference-indicative semantics, while observed user-item interactions may provide heterogeneous levels of preference evidence. Most existing methods treat content uncertainty and interaction uncertainty separately, which limits their ability to control the propagation of low-confidence signals throughout multimodal fusion, graph learning, and representation alignment. To address this issue, we propose a Multi-Level Contrastive Learning Framework with Reliability Estimation for Multimodal Recommendation (MRCL). First, we design a Dual-Prototype Residual Purification module, which constructs semantic and behavioral prototypes to refine multimodal representations by softly adjusting low-confidence feature components toward a fused prototype that integrates semantic neighborhood evidence and collaborative behavioral evidence. Second, we develop a cross-level reliability co-refinement mechanism to jointly refine item-level content reliability and edge-level interaction reliability, enabling reliability-aware graph propagation with soft confidence weights. Third, we introduce a Reliability-Stratified Cross-view Contrastive Learning objective that aligns semantic and collaborative views while emphasizing reliable anchors and medium-reliability semi-hard negatives. Extensive experiments on three public datasets demonstrate the superiority and robustness of MRCL over strong baselines.
Keywords:
Multimodal recommendation
Reliability-aware learning
Uncertainty modeling
Contrastive learning
Journal
K
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
Cited Papers
No cited papers available