Extreme Anomalous Oversampling Technique for Class Imbalance
Chittima Chiamanusorn, Krung Sinapiromsaran · 2017
Class imbalance problem is an important classification problem in machine learning which shows undesirable performance of a minority class. These minority instances have a tendency to be misclassified due to their tiny portion in a dataset. For a binary classification, they are labeled as positive while the rest are labeled as negative. This research proposes the novel parameter-free oversampling technique called the extreme anomalous oversampling technique, EXOT, based on an extreme anomalous score, EAS, and a negative anomalous score, NAS. This technique is used for rebalancing a data distribution that can be classified by any classifier. EAS of the instance p is the largest radius of an open ball centering at p containing only a single instance while NAS is the largest radius of an open ball centering at p without negative instances. EXOT synthesizes positive instances surrounding the original positive one using these two EAS and NAS. This work was conducted with three UCI datasets comparing SMOTE, borderline-SMOTE, safe-level SMOTE, and EXOT based on four classifiers which are C4.5, K-nearest neighbor classifier, multilayer perceptron, and naïve Bayes using precision, recall, F1-measure, and G-mean as performance measures. The results show the improvement of classification performances on a minority class on all three UCI datasets.