MUEnsemble: Multi-ratio Undersampling-Based Ensemble Framework for Imbalanced Data

Abstract

Class imbalance is commonly observed in real-world data, and it is still problematic in that it hurts classification performance due to biased supervision. Undersampling is one of the effective approaches to the class imbalance. The conventional undersampling-based approaches involve a single fixed sampling ratio. However, different sampling ratios have different preferences toward classes. In this paper, an undersampling-based ensemble framework, MUEnsemble, is proposed. This framework involves weak classifiers of different sampling ratios, and it allows for a flexible design for weighting weak classifiers in different sampling ratios. To demonstrate the principle of the design, in this paper, three quadratic weighting functions and a Gaussian weighting function are presented. To reduce the effort required by users in setting parameters, a grid search-based parameter estimation automates the parameter tuning. An experimental evaluation shows that MUEnsemble outperforms undersampling-based methods and oversampling-based state-of-the-art methods. Also, the evaluation showcases that the Gaussian weighting function is superior to the fundamental weighting functions. In addition, the parameter estimation predicted near-optimal parameters, and MUEnsemble with the estimated parameters outperforms the state-of-the-art methods.