Go to CVPR 2025 Workshop VisCon homepageSGBD: Sharpness-Aware Mirror Gradient with BLIP-Based Denoising for Robust Multimodal Product Recommendation
Keywords: Sharpness Aware Minimization, Recommendation System, Machine Learning, Deep Learning, Multimodal Model
Abstract: Multimodal recommender systems leverage diverse information to model user preferences and item features. While integrating multimodal data mitigates sparsity and cold-start challenges, it introduces information adjustment and noise risks. We analyze these systems through flat local minima and use BLIP's denoising capability to address inherent noise. Our proposed training strategy enhances model robustness during optimization. Through theoretical and empirical analyses, we demonstrate our approach's effectiveness across multiple recommendation models. The proposed Sharpness-Aware Mirror Gradient with BLIP-Based Denoising (SGBD) complements existing techniques and extends to advanced models, establishing a robust paradigm for multimodal recommendation.
Submission Number: 46
Loading