Improvement in the classification of data with class imbalance through a redistribution of classes by k-means

Abstract

In the field of machine learning, there are several challenges that affect the performance of classification algorithms. Some of these include the curse of dimensionality or class imbalance. The dimensionality curse is a phenomenon that occurs when the number of features (p) in a dataset increases significantly compared to the number of samples (n) available. On the other hand, class imbalance occurs when one or more classes in a dataset have significantly less representation than other classes. This decreases the performance of a classifier since it generates classification biases towards the majority class. Microarray data is widely used to analyze and understand gene expression on a global level. These provide information about the expression of thousands of genes simultaneously and can be used to classify different conditions or diseases. Such data exhibits both dimensionality curse and class imbalance complexities.

In this work, a method to divide the majority class into two or more classes by means of the k-means clustering algorithm in microarray datasets is presented. Classification is performed using a variety of state of the art classification algorithms. The proposed method exceeds the classification performance of the original methods as it is reported, taking into account the balanced accuracy and a 5-fold cross-validation. After performing the Mann-Whitney statistical test, it was determined that the proposed mehotd obtains significantly better results than when the original algorithms are used.