A novel data cleaning method for learning in imbalanced datasets based on k-nearest neighbors

Abstract

With the expansion of the applications of artificial intelligence and machine learning in various areas, many challenges have arisen in the training of learning algorithms. One of the most important challenges is the learning in imbalanced datasets. The imbalanced data generally refers to a classification problem where the number of data samples per class is not equally distributed. Typically there is a large amount of data for one class (referred to as the majority class) and much fewer data for the other class (referred to as the minority class). In such datasets, learning algorithms are biased toward learning the majority class to reach better accuracy, which leads to the lack of learning in minority class data. This article first introduces a method called neighbor competition scoring (NCS) for assigning scores to samples. Each data point is assigned a lower score if it is farther away from its class samples or closer to the samples of other classes. Then, with the help of these scores, neighbor competition undersampling (NCU) method is presented for undersampling the majority class samples that are less important than other samples. The proposed method has been compared with some popular data-based methods in 10 datasets and has performed better according to the experiments.