Robust Regression With Nonconvex Schatten P-Norm Minimization

Linear regression classification known as a classical supervised learning algorithm has been widely used in face recognition, image alignment, pose estimation and so on. Unfortunately, this algorithm often suffers from significant degradation in prediction accuracy when the presence of outlier or gross errors in the training data. To handle this issue, a novel robust regression is proposed by exploiting a nonconvex schatten p-norm minimization in this paper. Concretely, the l(p)-norm and nonconvex schatten p-norm are adopted to better approximate the l(0)-norm and rank minimization problem, respectively. Experimental results on several datasets demonstrate the superiority of our proposed model on classification, against to the state-of-the-art methods.