A Survey of mislabeled training data detection techniques for pattern classification

Pattern classification is an important part of machine learning. To use it, a classifier is trained on the training data and then predicts the label for the future unseen data. To obtain a classifier with good performance, the quality of the training data plays an important role. Unfortunately in many areas, it is difficult to provide absolutely clean data. This paper focuses on mislabeled data, which is one of the main types of noisy data. A number of mislabeled data detection techniques have been proposed; however, there is no survey work to summarize those techniques. This paper reviews the existing studies and classifies them into three types: Local learning-based, ensemble learning-based, and single learning-based methods. The technical details, advantages, and disadvantages of these methods are discussed.