Using fuzzy undersampling and fuzzy PCA to improve imbalanced classification through Rotation Forest algorithm

Abstract

This paper proposed a novel undersampling method to reduce the imbalance ratio of a dataset using fuzzy memberships degrees as well as utilizing a new fuzzy principal components analysis (F-PCA) for the classification through Rotation Forest algorithm. In the undersampling phase, first two membership functions are defined on each feature (dimension); one indicates the minority concept and the other shows majority concept. After that, each data sample receives a score based on its membership degrees in each dimension of the feature space. Majority samples with the highest scores are the best candidates of removal. Then during the Rotation Forest algorithm's train phase, a fuzzy Principal Component Analysis (F-PCA) is applied on the fuzzified values of samples which are produced in the undersampling phase. Moreover, these values are used to build the base classifiers of the ensemble. The obtained results illustrate the efficiency and noteworthy high performance of our proposed method comparing to the other state-of-the-art algorithms for class imbalance problem.