MahalCUSFilter: A Hybrid Undersampling Method to Improve the Minority Classification Rate of Imbalanced Datasets

Abstract

Class Imbalance problem has received considerable attention in the machine learning research. Among the methods which handle class imbalance problem, undersampling is a data level approach which preprocesses the data set to reduce the size of the majority class instances. Most of the existing undersampling methods apply either prototype selection or clustering techniques to balance the data set. They are effective and popular, but both processes are complex. Drawbacks of the cluster based undersampling methods are: The quality of the chosen majority class samples varies depending on clustering algorithm, number of clusters and also the convergence is difficult. Drawback of prototype selection methods is that they have to compare each majority instance with it’s k nearest neighbors to decide which majority class instance should be selected/discarded which is not only time consuming and is also difficult to implement for large datasets. Proposed undersampling method MahalanobisCentroidbasedUndersampingwithFilter (MahalCUSFilter) overcomes the above said problems: parameter dependence, complexity and information loss. Proposed method is used in conjunction with c4.5 and kNN classifiers, and found to improve the minority class classification rate of all datasets with comparable overall performance for the entire dataset. To the best of our knowledge this kind of grouping has not been used in undersampling to improve the classification accuracy of imbalanced data sets.