Hybrid sampling for imbalanced data

Abstract

Decision tree learning in the presence of imbalanced data is an issue of great practical importance, as such data is ubiquitous in a wide variety of application domains. We propose hybrid data sampling, which uses a combination of two sampling techniques such as random oversampling and random undersampling, to create a balanced dataset for use in the construction of decision tree classification models. The results demonstrate that our methodology is often able to improve the performance of a C4.5 decision tree learner in the context of imbalanced data.