Data quality is essential for the performance of deep neural networks in various fields. However, label noise and class imbalance are common data issues, which cause deep learning models to overfit in real-world scenarios. Recent research solves the learning with noisy labels (LNL) problem by employing label correction or loss adjustment methods, which often rely on uncertainty estimation. Unfortunately, these methods usually do not work well with imbalanced datasets. To address both noisy and imbalanced data biases, we analyze the limitations of current discriminative models in uncertainty estimation, and propose Class-aware Multi-granularity Co-Diffusion models (CaMCoD), which leverage a generative uncertainty and inconsistency loss adjustment method to generate labels more robustly. Specifically, we reframe the LNL problem as a robust diffusion-generative process, i.e., labels are generated by gradually refining an initial random guess. First, we use coarse-grained uncertainty from the diffusion model to achieve more accurate confidence estimates. This will guide the model to generate correct labels on a broader level. Then, we leverage the fine-grained inconsistency of co-diffusion models during reverse denoising to determine the learnable weight for each sample, which can mitigate the risk of the model overfitting to noisy samples. Finally, we apply class-aware loss adjustments to reduce data bias caused by class imbalance. Experiments on both synthetic and real-world datasets demonstrate that our method perform well in imbalanced and noisy scenarios. We provide our code on GitHub:https://github.com/SenyuHou/CaMCoD.

In image classification, the label quality of training data critically influences model generalization, especially for deep neural networks (DNNs). Traditionally, learning from noisy labels (LNL) can improve the generalization of DNNs through complex architectures or a series of robust techniques, but its performance improvement is limited by the discriminative paradigm. Unlike traditional ways, we resolve the LNL problems from the perspective of robust label generation, based on diffusion models within the generative paradigm. To expand the diffusion model into a robust classifier that explicitly accommodates more noise knowledge, we propose a Directional Label Diffusion (DLD) model. It disentangles the diffusion process into two paths, i.e., directional diffusion and random diffusion. Specifically, directional diffusion simulates the corruption of true labels into a directed noise distribution, prioritizing the removal of likely noise, whereas random diffusion introduces inherent randomness to support label recovery. This architecture enable DLD to gradually infer labels from an initial random state, interpretably diverging from the specified noise distribution. To adapt the model to diverse noisy environments, we design a low-cost label pre-correction method that automatically supplies more accurate label information to the diffusion model, without requiring manual intervention or additional iterations. Our approach outperforms state-of-the-art methods on both simulated and real-world noisy datasets. Code is available at https://github.com/SenyuHou/DLD.

The linear regression model is simple in form and easy to estimate; nevertheless, irrelevant features will raise the difficulty of its tasks. Feature selection is generally adopted to improve the model performance. Unfortunately, traditional regression feature selection methods may not work for data with noise or outliers. Although some robust methods for certain specific error distributions have been proposed, they may not perform well because the distribution of representation error is often unknown for real data. This paper proposes a regression feature selection method for unknown noise named Mixture of Gaussians LASSO (MoG-LASSO), in which feature selection and model training will be achieved simultaneously. MoG is adopted to model unknown noises, and M-estimation is used to acquire the weighted squared error loss. By alternatively and iteratively updating the regression coefficient and parameters of MoG, the influence of unknown noise can be reduced effectively. Furthermore, MoG-LASSO achieves feature selection by the$L_{1}$regularization term, which can further improve the performance of the model. Experimental results on artificial data and benchmark data sets demonstrate that MoG-LASSO has better robustness and sparsity for data sets with irrelevant features. Additionally, experimental results on face recognition databases show the performance advantage of MoG-LASSO over state-of-the-art methods in the presence of illumination variations.