K-Neighbor over-sampling with cleaning data: a new approach to improve classification performance in data sets with class imbalance

摘要

The problems of class imbalance have attracted concerns of researchers in the last few years. Class imbalance problems occur when the data has unbalanced proportions between groups or levels of the response variable. The class/group with a dominant proportion is called as the majority class, while the small proportion class is known as the minority. These problems relate to the creation of bias in parameter estimation in a parametric model such as logistic regression and high misclassification rate for the minority class. Eventually, it might create a risks in the policymaking. To overcome these problems, a new approach called K-Neighbor Over-sampling (KNOS) with cleaning data is proposed in this paper. Unlike the other methods such as SMOTE, Border-line SMOTE (BLS), and Safe Level-SMOTE (SLS), in generating synthetic data, KNOS employs K minority class observations, then KNOS proceeds further by removing some majority class observations. Notice that, similar to BLS and SLS, KNOS only generates synthetic samples from original samples which are safe. In this paper KNOS has been applied to logistic regression classification method. The results showed that KNOS method produced a higher performance in terms of AUC,G-Mean, and sensitivity compared to BLS and SLS. Moreover, our study has also shown that KNOS produced more consistent result than the other approaches.