Range-limited Augmentation for Few-shot Learning in Tabular Data

Kyungeun Lee, Moonjung Eo, Hye-Seung Cho, Suhee Yoon, Sanghyu Yoon, Ye Seul Sim, Woohyung Lim

Keywords: Tabular representation learning; Contrastive learning; Few-shot learning; Data augmentation

TL;DR: We propose range-limited augmentation for contrastive learning in tabular data, outperforming existing few-shot learning techniques by preserving task-relevant information across a benchmark of 42 datasets and 31 algorithms.

Abstract: Few-shot learning is essential in many applications, particularly in tabular domains where the high cost of labeling often limits the availability of annotated data. To address this challenge, we propose range-limited augmentation for contrastive learning in tabular domains. Our augmentation method shuffles or samples values within predefined feature-specific ranges, preserving semantic consistency during contrastive learning to enhance few-shot classification performance. To evaluate the effectiveness of our approach, we introduce FeSTa (Few-Shot Tabular classification benchmark), a benchmark consisting of 42 tabular datasets and 31 algorithms. On this benchmark, contrastive learning with our augmentation method effectively preserves task-relevant information and significantly outperforms existing approaches, including supervised, unsupervised, self-supervised, semi-supervised, and foundation models. In particular, our method achieves an average rank of 2.3 out of 31 algorithms in the 1-shot learning scenario, demonstrating its robustness and effectiveness when labeled data is highly limited. The benchmark code is available in the supplementary material.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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: 9039