A Noisy-sample-removed Under-sampling Scheme for Imbalanced Classification of Public Datasets

Abstract

Abstract Classification technology plays an important role in machine learning. In the process of classification, the presence of noisy samples in datasets tends to reduce the performance of a classifier. This work proposes a clustering-based Noisy-sample-Removed Under-sampling Scheme (NUS) for imbalanced classification. First, the samples in the minority class are clustered. For each cluster, its center is taken as a spherical center, and the distance of the minority class samples farthest from the cluster center is taken as the radius to form a hypersphere. The Euclidean distance from the center of the cluster to every of the majority samples is calculated to decide if they are in the hypersphere. Then, we propose a NUS-based policy to decide if a majority sample in the hypersphere is a noisy sample. Similarly, the noises samples of the minority class are found. Second, We remove noisy-samples from the majority and minority classes and propose NUS. Finally, logistics regression, Decision Tree, and Random Forest are used in NUS as the base classifiers, respectively and compare with Random Under-Sampling (RUS), EasyEnsemble (EE), and Inverse Random Under-Sampling (IRUS) on 13 public datasets. Results show that our method can improve the classification performance in comparison with its state-of-the art peers.