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

Mingming Zhang, Zhiqing Xiao, NINGTAO WANG, Xing Fu, Yu Cheng, Weiqiang Wang, Junbo Zhao

Keywords: test time adaptation; transfer learning; representation learning

Abstract: Tabular data is ubiquitous across real-world applications. While self-supervised learning has advanced representation learning for tabular data, most methods assume the unrealistic IID setting. In practice, tabular data often exhibits distribution shifts, including both label and covariate shifts, rendering existing domain generalization or test-time adaptation techniques from computer vision ineffective. To address this, 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). It leverages high-confidence and domain-consistent pseudo-labels to estimate and correct for target label distribution shifts. Subsequently, it employs self-training and entropy minimization, guided by these confident samples, to adapt the model to the out-of-distribution test data. Extensive experiments across diverse distribution shift scenarios demonstrate that OT3A significantly outperforms existing methods, highlighting its efficacy and practicality for adapting to real-world tabular data.

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

Submission Number: 12559