Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AdapTable: Test-Time Adaptation for Tabular Data via Shift-Aware Uncertainty Calibrator and Label Distribution Handler

Changhun Kim, Taewon Kim, Seungyeon Woo, June Yong Yang, Eunho Yang

TL;DR

The paper tackles distribution shifts in tabular data, which undermine reliability in real-world deployments. It proposes AdapTable, a two-stage test-time adaptation framework leveraging a shift-aware uncertainty calibrator (via a graph neural network over table columns) and a label distribution handler to align predictions with the target label distribution without updating model parameters. Theoretical analysis bounds the adaptation error and highlights the importance of accurately estimating the target distribution and controlling representation shift. Empirically, AdapTable achieves state-of-the-art performance across six datasets and six corruptions, with substantial gains on HELOC and strong cross-architecture robustness, making it a practical, privacy-friendly TTA solution for tabular domains.

Abstract

In real-world scenarios, tabular data often suffer from distribution shifts that threaten the performance of machine learning models. Despite its prevalence and importance, handling distribution shifts in the tabular domain remains underexplored due to the inherent challenges within the tabular data itself. In this sense, test-time adaptation (TTA) offers a promising solution by adapting models to target data without accessing source data, crucial for privacy-sensitive tabular domains. However, existing TTA methods either 1) overlook the nature of tabular distribution shifts, often involving label distribution shifts, or 2) impose architectural constraints on the model, leading to a lack of applicability. To this end, we propose AdapTable, a novel TTA framework for tabular data. AdapTable operates in two stages: 1) calibrating model predictions using a shift-aware uncertainty calibrator, and 2) adjusting these predictions to match the target label distribution with a label distribution handler. We validate the effectiveness of AdapTable through theoretical analysis and extensive experiments on various distribution shift scenarios. Our results demonstrate AdapTable's ability to handle various real-world distribution shifts, achieving up to a 16% improvement on the HELOC dataset.

AdapTable: Test-Time Adaptation for Tabular Data via Shift-Aware Uncertainty Calibrator and Label Distribution Handler

TL;DR

The paper tackles distribution shifts in tabular data, which undermine reliability in real-world deployments. It proposes AdapTable, a two-stage test-time adaptation framework leveraging a shift-aware uncertainty calibrator (via a graph neural network over table columns) and a label distribution handler to align predictions with the target label distribution without updating model parameters. Theoretical analysis bounds the adaptation error and highlights the importance of accurately estimating the target distribution and controlling representation shift. Empirically, AdapTable achieves state-of-the-art performance across six datasets and six corruptions, with substantial gains on HELOC and strong cross-architecture robustness, making it a practical, privacy-friendly TTA solution for tabular domains.

Abstract

In real-world scenarios, tabular data often suffer from distribution shifts that threaten the performance of machine learning models. Despite its prevalence and importance, handling distribution shifts in the tabular domain remains underexplored due to the inherent challenges within the tabular data itself. In this sense, test-time adaptation (TTA) offers a promising solution by adapting models to target data without accessing source data, crucial for privacy-sensitive tabular domains. However, existing TTA methods either 1) overlook the nature of tabular distribution shifts, often involving label distribution shifts, or 2) impose architectural constraints on the model, leading to a lack of applicability. To this end, we propose AdapTable, a novel TTA framework for tabular data. AdapTable operates in two stages: 1) calibrating model predictions using a shift-aware uncertainty calibrator, and 2) adjusting these predictions to match the target label distribution with a label distribution handler. We validate the effectiveness of AdapTable through theoretical analysis and extensive experiments on various distribution shift scenarios. Our results demonstrate AdapTable's ability to handle various real-world distribution shifts, achieving up to a 16% improvement on the HELOC dataset.
Paper Structure (58 sections, 1 theorem, 29 equations, 12 figures, 9 tables)

This paper contains 58 sections, 1 theorem, 29 equations, 12 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Analysis of Tabular Distribution Shifts
  3. Indistinguishable Representations
  4. Importance of Label Distribution Shifts
  5. AdapTable
  6. Test-Time Adaptation Setup for Tabular Data
  7. Shift-Aware Uncertainty Calibrator
  8. Label Distribution Handler
  9. Theoretical Insights
  10. Experiments
  11. Experimental Setup
  12. Datasets.
  13. Model architectures and baselines.
  14. Evaluation metrics and implementation details.
  15. Main Results
  16. ...and 43 more sections

Key Result

Theorem 3.1

Let $\hat{Y}|X$ and $\hat{Y}_{o}|X$ be defined as follows: Given the error $\epsilon(\hat{Y} | X) = \mathbb{P}(\hat{Y} \ne Y | X)$, with true labels $Y$ of inputs $X$, the error gap $| \epsilon(\hat{Y} | X_s) - \epsilon(\hat{Y}_{o}|X_t) |$ is upper bounded by where $K_1$ and $K_2$ are constants related to $p_t(y)$, and $p_s(y)$, respectively.

Figures (12)

  • Figure 1: Latent space visualization with t-SNE comparing (a) tabular data tableshift and (b) image data openml_benchmark. Reliability diagrams of (c) underconfident and (d) overconfident scenarios are shown. All experiments are conducted using an MLP architecture.
  • Figure 2: Label distribution of (a) source domain, (b) target domain, (c) estimated label distribution using pseudo labels, and (d) corrected label distribution of AdapTable are shown using MLP on HELOC dataset.
  • Figure 3: The overall pipeline of the AdapTable framework. AdapTable employs a per-sample temperature scaling to correct overconfident predictions by treating each column as a graph node, enabling a shift-aware uncertainty calibrator with graph neural networks to capture both individual and complex feature shifts (Section \ref{['subsec:calibrator']}). It also estimates the label distribution of the current test batch and adjusts the model’s predictions accordingly (Section \ref{['subsec:label_distribution_handler']}).
  • Figure 4: The average macro F1 score for AdapTable and TTA baselines is reported under six common corruptions using MLP across three datasets within the TableShift tableshift benchmark.
  • Figure 5: Ablation study on the shift-aware uncertainty calibrator using MLP for the HELOC dataset. (a) and (b) show reliability diagrams before and after calibration, while (c) depicts the average temperature relative to the maximum mean discrepancy (MMD) between the training set and the sampled test sets.
  • ...and 7 more figures

Theorems & Definitions (3)

  • Theorem 3.1
  • Definition C.1
  • proof