Browse Teach Curate

Diffusion-Based Contrastive Learning with Semantic Alignment for Multimodal Recommendation

DiffCL uses a diffusion model to generate denoised contrastive views for self-supervised contrastive learning in multimodal recommendation, aligns each item's visual and textual semantics through stable ID embeddings, and introduces an Item-Item Graph to enrich representations under data sparsity, jointly improving recommendation accuracy from noisy multimodal item features.

Rosie Text

DiffCL: A Diffusion-Based Contrastive Learning Framework with Semantic Alignment for Multimodal

Introduces DiffCL, a framework for multimodal recommendation (recommenders that use items' image and text features, not just user-item interactions). It targets three problems: noise in multimodal fe…