Optimized Oversampling
Keywords: Machine Learning, Imbalanced Datasets, Optimization
Abstract: Many classification problems that arise in practice feature imbalanced datasets, a regime in which a lot of machine learning (ML) models show diminished performance. To address class imbalance, techniques like undersampling and oversampling are used to improve the model's performance. In this paper, we introduce a new oversampling framework, Optimized Oversampling ($O^{2}$), which generates synthetic minority class points by maximizing the probability of belonging to the minority class, which is estimated by a trained classification model. We show theoretically, under mild assumptions, that the points generated by $O^{2}$ are more likely to belong to the minority class than those generated by other approaches. Further, we benchmark $O^{2}$ against state-of-the-art oversampling methods on 16 publicly available imbalanced datasets using Classification Trees (CART) and Logistic Regression (LR) for the downstream classification task. The numerical experiments show that $O^{2}$ has an edge over current state-of-the-art oversampling methods, which is more pronounced on CART.
Supplementary Material: zip
Primary Area: optimization
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8008
Loading