An encapsulated framework for imbalanced classification based on multiratio undersampling

Abstract

Real-world data frequently exhibits class imbalance, which is troublesome since it impairs
classification ability because of biased supervision. One efficient resampling technique for the class imbalance is
undersampling. One fixed sample ratio is used in the traditional undersampling-based methods. Nonetheless,
preferences for classes vary depending on the sampling ratio. This study proposes MUEnsemble, an ensemble
architecture based on undersampling. A versatile design for weighting weak classifiers in various sampling ratios
is made possible by this framework, which incorporates weak classifiers of various sampling ratios. This paper
presents a uniform weighting function and a Gaussian weighting function to illustrate the design principle.
According to a thorough experimental review, MUEnsemble performs better than state-of-the-art techniques that
rely on undersampling or oversampling in terms of recall, gmean, F-measure, and ROC-AUC metrics.
Furthermore, the evaluation demonstrates the superiority of the Gaussian weighting function over the uniform
weighting function. This suggests that the various preferences of sampling ratios toward classes can be captured
by the Gaussian weighting function. An analysis of the impacts of the Gaussian weighting function's parameters
reveals that recall can be used to determine the function's parameters, which is desirable in many real-world
applications.