Online Test-Time Adaptation in Tabular Data with Minimal High-Certainty Samples

Tabular data is ubiquitous across real-world applications, yet most representation learning methods still assume an unrealistic IID setting. In practice, tabular streams often exhibit mixed covariate and label shifts, making existing domain generalization or vision-oriented test-time adaptation methods ineffective. We propose a simple yet effective \textbf{O}nline \textbf{T}est-\textbf{T}ime \textbf{A}daptation approach for \textbf{T}abular data (OT3A). OT3A uses high-confidence and domain-consistent pseudo-labels to estimate and correct target label distribution shifts, then applies self-training and entropy minimization on these reliable samples for online adaptation. Extensive experiments across diverse distribution shift scenarios show that OT3A significantly outperforms existing methods, highlighting its efficacy and practicality for real-world tabular adaptation. \textit{Our code will be open to the public in the future.