Testing Uncertainty-Aware Conditional Representation Consistency for Tabular Counterfactuals

Tabular data are widely used in healthcare, finance, science, and other decision-making applications, yet robust representation learning remains challenging because small feature perturbations may correspond either to meaningful variations or to unrealistic samples. Standard consistency regularization treats such perturbations uniformly and can therefore enforce invariance when a prediction should change. This paper introduces CONSENT, an uncertainty-aware conditional representation learning framework that applies consistency regularization only when a perturbed sample satisfies three criteria: a distance-based validity score, high predictive confidence, and prediction preservation. The framework addresses the need for a more selective consistency mechanism that accounts for the validity and reliability of local perturbations rather than enforcing representation invariance indiscriminately. It provides a unified formulation for incorporating counterfactual validity, predictive uncertainty, and prediction preservation into representation consistency for tabular data. Experiments on three tabular datasets examine the effectiveness of conditional representation consistency under local perturbations, providing empirical evidence on the potential and limitations of uncertainty-aware consistency mechanisms for robust tabular representation learning.