Improving imbalanced learning through a heuristic oversampling method based on k-means and SMOTE

摘要

Learning from class-imbalanced data continues to be a common and challengingproblem in supervised learning as standard classification algorithms aredesigned to handle balanced class distributions. While different strategiesexist to tackle this problem, methods which generate artificial data to achievea balanced class distribution are more versatile than modifications to theclassification algorithm. Such techniques, called oversamplers, modify thetraining data, allowing any classifier to be used with class-imbalanceddatasets. Many algorithms have been proposed for this task, but most arecomplex and tend to generate unnecessary noise. This work presents a simple andeffective oversampling method based on k-means clustering and SMOTEoversampling, which avoids the generation of noise and effectively overcomesimbalances between and within classes. Empirical results of extensiveexperiments with 71 datasets show that training data oversampled with theproposed method improves classification results. Moreover, k-means SMOTEconsistently outperforms other popular oversampling methods. An implementationis made available in the python programming language.