Simulating Complexity Measures on Imbalanced Datasets

Abstract

Classification tasks using imbalanced datasets are not challenging on their own. Classification models perform poorly on the minority class when the datasets present other difficulties, such as class overlap and complex decision border. Data complexity measures can identify such difficulties, better dealing with imbalanced datasets. They can capture information about data overlapping, neighborhood, and linearity. Even though they were recently decomposed by classes to deal with imbalanced datasets, their high computational cost prevents their use on applications with a time restriction, such as recommendation systems or high dimensional datasets. In this paper, we use a Meta-Learning approach to estimate the decomposed data complexity measures. We show that the simulated measures assess the difficulty of the dataset after applying preprocessing techniques to different sample sizes. We also show that this approach is significantly faster than computing the original measures, with a statistically similar estimation error for both classes.