UA-Tab: Uncertainty-aware self-supervised representation learning for tabular data

Tabular data, a cornerstone of real-world applications in domains such as healthcare, finance, and infrastructure, pose unique challenges for deep learning due to their heterogeneous feature types, lack of spatial structure, and frequent presence of noisy or missing values. Although recent advances in self-supervised learning (SSL) have enabled progress in tabular representation learning, existing methods often overlook a critical aspect of real-world data—feature-level uncertainty. In this paper, we propose UA-Tab (Uncertainty-Aware self-supervised representation learning for Tabular data), a novel self-supervised framework that explicitly models and leverages uncertainty in tabular data. UA-Tab introduces two key innovations: (1) a unified attention-uncertainty mechanism that identifies and emphasizes informative yet reliable features while down-weighting noisy or ambiguous inputs, and (2) a latent-space perturbation strategy for view generation, enabling contrastive learning without the need for unreliable negative sampling. Through comprehensive experiments on the Adult Income dataset under a wide range of synthetic noise scenarios—including value corruption, masking, mode collapse, and out-of-distribution anomalies—UA-Tab demonstrates superior robustness and generalization compared to existing SSL baselines such as VIME, SCARF, STaB, SubTab, and TabDeCo. Notably, UA-Tab consistently achieves the highest accuracy and macro F1-score across conditions while maintaining interpretability through its uncertainty outputs. These results highlight UA-Tab’s potential to serve as a reliable and transparent foundation for self-supervised tabular learning, especially in environments where data quality varies and explainability is critical.