Combating Sub-Clusters Effect in Imbalanced Classification

Abstract

Approaches to imbalanced classification problem usually focus on rebalancing the class sizes, neglecting the effect of hidden structure within the majority class. The aim of this paper is to highlight the effect of sub-clusters within the majority class on detecting minority class instances, and handle imbalanced classification by learning the structure in the data. We propose a decomposition based approach to two-class imbalanced classification problem. This approach works by first learning the hidden structure of the majority class using an unsupervised learning algorithm. Thus, transforming the classification problem into several classification sub-problems. The base classifier is constructed on each sub-problem. The ensemble is tuned to increase its sensitivity towards minority class. The proposed approach overcomes the limitations of conventional classifiers on imbalanced problem, and combats imbalance by learning hidden structure in the majority class, which is neglected in most existing works. We demonstrate the performance of the proposed approach on several datasets.