Tabular Deep-SMOTE: A supervised autoencoder-based minority-oversampling technique for class-imbalanced tabular classification

Daniel Lawton, Moshe Salhov, Ofir Lindenbaum, Amir Averbuch

Keywords: supervised learning, tabular data, imbalanced datasets, classification, minority oversampling

TL;DR: Proposal of a new minority oversampling technique for imblanced tabular datasets

Abstract: Class imbalance, present in many real-world tabular datasets, may cause machine-learning models to under-classify minority samples, which are often highly significant. This work proposes a new oversampling method called Tabular Deep-SMOTE (TD-SMOTE), which harnesses the class labels to improve synthetic sample generation via autoencoders. The method is based on oversampling in an alternative space shaped by a metric-learning loss. Such spaces tend to be more semantic and obtain higher class separation and density, which improves the quality of samples generated by linear interpolations over the observed minority samples. In addition, we propose a synthetic samples filtering scheme based on the decision boundary of a pre-trained tabular classifier to guarantee the quality of synthetic samples. Compared to common and leading oversampling methods, the method achieves improved classification performance in an extensive set of experiments that includes over 36 publicly available datasets.

Primary Area: general machine learning (i.e., none of the above)

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/2024/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: 2806