Fuzzy rough label modification learning for unlabeled and mislabeled data

Abstract

Mislabeling is one of the major challenges in semi-supervised learning methods. Most existing approaches based on fuzzy rough sets typically assume that labeled data is accurate and free from errors. This assumption often overlooks the presence of incorrect labels, which can weaken the generalization ability and reduce the robustness of the learning algorithms. To address these shortcomings, we propose a new method for label modification based on fuzzy rough sets, called the Fuzzy Rough Set Label Modification Filter (RSLMF), which is designed to handle both unlabeled and mislabeled data. Specifically, the proposed RSLMF consists of two main steps: detection of mislabeled samples and their subsequent correction. The filter employs the theoretical framework of fuzzy rough sets to identify mislabeled samples in data and subsequently correct their labels. For unlabeled samples, their potential labels are inferred by analyzing their correlation with labeled data through fuzzy rough sets. Additionally, the topological structures of the corrected sample set and the boundary sample set are thoroughly explored during the label modification process. Experimental results demonstrate that the proposed method can accurately modify the labels of mislabeled samples and effectively suppress noise.