Robust Semi-supervised Classification Method Search For Noisy Labels Based On Self-paced Learning

Data Labeling is a cumbersome,time-consuming task easily prone to error.On one hand,it makes the number of labeled data much limited,while there is more unlabeled data.On the other hand,it generates noisy labels in the process of data labeling,but many machine learning algorithms are sensitive to noisy labels.So people hope to search a stable algorithm that can utilize a large amount of unlabeled data and be robust to noisy label data.For this problem,we propose a semi-supervised classification framework called Self-Paced Manifold Regularization that is robust to noisy labels.This framework integrates self-paced learning regime into the Manifold Regularization framework to select labeled training data in a theoretically sound manner,meanwhile,it utilize the discrimination information contained in the sparse coding to control smoothness of classifier.Finally,the alternative search strategy is adopted to deal with the optimization problem,and the framework obtains an explicit multi-class classifier.The classifier is robust to noisy labels,and take into account of the complexity and smoothness of the classifier,to reduce generalization error.Experimental results demonstrate the effectiveness of the proposed method.