Robust Multi-View Multi-Label Learning via Uncertainty-Gated Diffusion and Dynamic Graph mRMR

Incomplete Multi-view Noisy Multi-label Learning (IMvNML), which involves both missing views and noisy labels, has attracted increasing attention in real-world applications. Existing IMvNML methods have made progress but still face two key issues: (i) They mainly focus on deterministic view completion but overlook the inherent generation uncertainty, leading to the introduction of unreliable reconstructions into latent representations. (ii) They primarily utilize static topological structures for label correction but overlook the dynamic feedback from refined features, preventing the joint improvement of feature robustness and label reliability. To address these issues, we propose a unified framework, named U-DMR, to establish a virtuous cycle via uncertainty-gated feature selection and closed-loop label refinement. Specifically, we introduce a Fast Adversarial Diffusion (FAD) method integrated with an Uncertainty-Gated mRMR (UG-mRMR) strategy to provide reliability-aware view completion and robust feature selection by reducing the influence of unreliable imputed features. Then, we design a Dynamic Graph Label Refinement (DGLR) method to improve label reliability by leveraging the robust features to build an evolving semantic graph. The refined labels provide cleaner supervision, which in turn iteratively improves feature learning. Extensive experiments on seven benchmark datasets demonstrate that U-DMR consistently outperforms state-of-the-art methods. The code and datasets are available at https://anonymous.4open.science/r/U-DMR-2538/.