Novel fuzzy clustering-based undersampling framework for class imbalance problem

The class imbalance problem occurs in various real-world datasets. Although it is considered that samples of the classes of a dataset are evenly distributed, in many cases, datasets are highly imbalanced. Classification of such datasets is challenging in machine learning. Researchers have developed many approaches to solve the class imbalance problem, such as resampling and ensemble methods. In resampling methods, minority class samples are increased (oversampling), or majority class samples are reduced (under-sampling). In contrast, the ensemble methods classify various subsets of data where classification results are combined to provide the final result. The authors have introduced a new fuzzy C-mean clustering-based under-sampling method in the present study. We performed experiments using newly proposed method over 30 small-scale imbalanced datasets. The results obtained revealed that the proposed method improves the classification performance. The average sensitivity improved by 1% and the average balance accuracy improved by 3% as compared to k-means undersampling method. The results of this study would be useful in classification of imbalanced datasets of various domains.