Enhancing Random Oversampling for Imbalanced Classification

Abstract

In classification problems, imbalanced data is a common challenge, as classifiers often exhibit a tendency to assign new sample points to the majority class. This leads to suboptimal prediction performance for the minority class. Therefore, it is imperative to mitigate the imbalanced data problem. Random oversampling is a simple employed technique to address class imbalance in datasets. Nevertheless, a subset of the sampled data points may prove inconsequential for the classification process. This research introduces a novel approach: the Hotelling Important Data Point Oversampling Algorithm (HIDPO), an improved version of Random Oversampling. This study aimed to compare classification performance using original data and data from oversampling techniques between Random Oversampling method and the proposed HIDPO method across 96 simulated scenarios. These scenarios varied in four parameters: 1) imbalance rate (IR), 2) the number of relevant predictor variables (RelVar), 3) the difference in means of predictor variables between the minority and majority groups (ClassDif), and 4) sample size (n). Logistic regression models were employed to perform classification tasks. The empirical findings demonstrated that the HIDPO method yields the highest F-measure in scenarios with minimal differences between the minority and majority classes, particularly in cases of severe imbalance, which present challenges for classification. Regarding true positive rate and true negative rate, the HIDPO method yielded moderate values.