Joint Behavior-guided and Modality-coherence Conditional Graph Diffusion Denoising for Multi Modal Recommendation

A new research paper introduces JBM-Diff (Joint Behavior-guided and Modality-coherence Conditional Graph Diffusion Denoising), a novel AI architecture designed to solve core problems plaguing modern recommendation systems. Built by researchers Xiangchen Pan and Wei Wei, the model specifically targets the 'noise' inherent in user data: false clicks from accidental taps and irrelevant visual features from product images that don't reflect true user preference. Traditional models often misinterpret this noise, leading to poor suggestions.

JBM-Diff employs a two-pronged, graph-based diffusion approach. First, it uses a conditional diffusion model to denoise multimodal features (like text, images, and video), stripping away information unrelated to user preferences. Second, it analyzes user behavior sequences to identify and down-weight unreliable feedback, such as accidental clicks. The system then uses multi-view message propagation to better align the cleaned collaborative signals with the cleaned item features. Extensive testing on three public datasets demonstrates that this joint denoising strategy significantly boosts ranking accuracy compared to previous Graph Convolutional Network (GCN) methods.