A Diversity-Based Synthetic Oversampling Using Clustering for Handling Extreme Imbalance

Abstract

Imbalanced data are typically observed in many real-life classification problems. However, mainstream machine learning algorithms are mostly designed with the underlying assumption of a relatively well-balanced distribution of classes. The mismatch between reality and algorithm assumption results in a deterioration of classification performance. One form of approach to address this problem is through re-sampling methods, although its effectiveness is limited; most re-sampling methods fail to consider the distribution of minority and majority instances and the diversity within synthetically generated data. Diversity becomes increasingly important when minority data becomes more sparse, as each data point becomes more valuable. They should all be considered during the generation process instead of being regarded as noise. In this paper, we propose a cluster-based diversity re-sampling method, combined with NOAH algorithm. Neighbourhood-based Clustering Diversity Over-sampling (NBCDO) is introduced with the aim to complement our previous cluster-based diversity algorithm Density-based Clustering Diversity Over-sampling (DBCDO). It first uses a neighbourhood-based clustering algorithm to consider the distribution of both minority and majority class instances, before applying NOAH algorithm to encourage diversity optimisation during the generation of synthetic instances. We demonstrate the implementation of both cluster-based diversity methods by conducting experiments over 10 real-life datasets with ≤ 5% imbalance ratio and show that our proposed cluster-based diversity algorithm (NBCDO, DBCDO) brings performance improvements over its comparable methods (DB-SMOTE, MAHAKIL, KMEANS-SMOTE, MC-SMOTE).