Revisiting Class Imbalance: A Generalized Notion for Oversampling

Class imbalance is a salient problem in both machine learning and data mining realms. Sampling techniques have become a cornerstone in solving this challenge, as they enable the creation of class-balanced datasets that is essential for robust model training. Addressing class imbalance not only enhances the predictive accuracy of machine learning algorithms but also ensures fair and unbiased decision-making across various applications, making it a critical aspect of research and development in various sectors. Through this work, we introduce the concept of a generalized oversampling function, unifying existing synthetic oversampling approaches. We explore diverse design decisions for such a function, presenting six functions categorized as linear and non-linear variants. We provide extensive experiments with these functions to gain an in-depth understanding of their behavior. Through our experiments, we observe that the best-performing function is primarily data-driven. Also, it is perceived that non-linear functions like minimum and maximum often depict higher learning capacity and steady performance in comparison to their linear counterparts mainly due to their ability in modelling non-trivial patterns. While moderate input counts would yield desirable performance in these functions, we can see varying robustness from these functions for distorted input data.