Adaptive neighbor synthetic minority oversampling technique under 1NN outcast handling

Abstract

SMOTE is an effective oversampling technique for a class imbalance problem due to its simplicity and relatively high recall value. One drawback of SMOTE is a requirement of the number of nearest neighbors as a key parameter to synthesize instances. This paper introduces a new adaptive algorithm called Adaptive neighbor Synthetic Minority Oversampling Technique (ANS) to dynamically adapt the number of neighbors needed for oversampling around different minority regions. This technique also defines a minority outcast as a minority instance having no minority class neighbors. Minority outcasts are neglected by most oversampling techniques but instead, an additional outcast handling method is proposed for the performance improvement via a 1-nearest neighbor model. Based on our experiments in UCI and PROMISE datasets, generated datasets from this technique have improved the accuracy performance of a classification, and the improvement can be verified statistically by the Wilcoxon signed-rank test.

Keywords

• class imbalance problem
• oversampling
• SMOTE
• adaptive neighbors approach
• minority outcast