Research map / Label-efficient and robust learning / Test-time adaptation
Fully Test-time Adaptation for Tabular Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FtaT, a fully test-time adaptation method for tabular data that combines label-distribution tracking, neighborhood-consistency weighting, and an online ensemble over learning rates to handle label and covariate shifts.
Problem
Fully test-time adaptation of deep tabular models under covariate and label distribution shifts, where existing image-oriented FTTA methods fail because augmentations are ineffective and adaptation is sensitive to task, model and learning rate.
Approach
Replaces image augmentation consistency with neighborhood consistency among batch samples, estimates the shifted label distribution from confident predictions with a covariate-shift correction and temporal smoothing, and ensembles several online models trained with different learning rates weighted by batch loss.
Key results
- On MLP backbone, FtaT reaches 64.09% accuracy on HELOC versus 54.37% for non-adaptation, and 66.77% average accuracy across six datasets versus 62.45% for non-adaptation.
- On DIABETE with MLP, FtaT achieves 61.66% accuracy and 59.27% F1, versus 60.81% and 51.18% for non-adaptation.
- Baselines TENT, LAME, ODS and others often fall below the non-adaptation baseline on tabular shifts, e.g. LAME drops to 43.10% accuracy on HELOC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FtaT instead of the strongest listed baseline | MLP · Average across benchmarks | accuracy: 61.59 -> 66.77 | Helped |
| Use FtaT instead of the strongest listed baseline | MLP · Average across benchmarks | balanced accuracy: 63.78 -> 64.96 | Helped |
| Use FtaT instead of the strongest listed baseline | MLP · Average across benchmarks | F1: 60.57 -> 72.00 | Helped |
| Use FtaT instead of the strongest listed baseline | TabTransformer · Average across benchmarks | accuracy: 60.37 -> 66.14 | Helped |
| Use FtaT instead of the strongest listed baseline | TabTransformer · Average across benchmarks | balanced accuracy: 62.82 -> 64.40 | Helped |
| Use FtaT instead of the strongest listed baseline | TabTransformer · Average across benchmarks | F1: 60.09 -> 69.03 | Helped |
| Use FtaT instead of the strongest listed baseline | FT-Transformer · Average across benchmarks | accuracy: 59.69 -> 64.01 | Helped |
| Use FtaT instead of the strongest listed baseline | FT-Transformer · Average across benchmarks | balanced accuracy: 62.49 -> 62.54 | Helped |
| Use FtaT instead of the strongest listed baseline | FT-Transformer · Average across benchmarks | F1: 57.04 -> 69.56 | Helped |
| Add Confident Distribution Optimizer (ablation) | FtaT · DIABETE | F1: 51.26±1.69 -> 59.27±0.96 | Helped |
| Add Confident Distribution Optimizer (ablation) | FtaT · HELOC | accuracy: 62.73±0.23 -> 64.09±1.14 | Helped |
| Add Local Consistent Weighter (ablation) | FtaT · DIABETE | F1: 55.61±1.67 -> 59.27±0.96 | Helped |
Surprising findings
- Standard FTTA methods (TENT, LAME, ODS, CoTTA with perturbation augmentation) frequently underperform the non-adaptation baseline on shifted tabular data.
- Increasing perturbation augmentation strength in CoTTA steadily degrades accuracy on tabular data.
Limitations
- The method lacks deep theoretical understanding, as acknowledged by the authors.
- Gains are inconsistent: on some datasets (e.g., ANES accuracy, Health Ins. balanced accuracy) FtaT is competitive rather than best.
- Hyperparameter choices (alpha, epsilon, beta, learning rates) are tuned per setting, and results are averaged over only three seeds.
Experimental setup
Datasets: HELOC ANES ASSISTMENTS DIABETES Hypertension Health Ins.
Benchmarks: HELOC ANES ASSISTMENTS DIABETES Hypertension Health Ins.
Metrics: accuracy balanced accuracy F1
Baselines: Non-Adaptation TENT EATA LAME CoTTA ODS SAR
Builds on: TENT CoTTA ODS LAME EATA
Closest papers by meaning
- TabLog: Test-Time Adaptation for Tabular Data Using Logic Rules (ICML 2024 · similarity 0.91)
- TabDPT: Scaling Tabular Foundation Models on Real Data (NeurIPS 2025 · similarity 0.85)
- TabFSBench: Tabular Benchmark for Feature Shifts in Open Environments (ICML 2025 · similarity 0.85)
- TEA: Test-time Energy Adaptation (CVPR 2024 · similarity 0.85)
- Large Scale Transfer Learning for Tabular Data via Language Modeling (NeurIPS 2024 · similarity 0.85)
- ConTextTab: A Semantics-Aware Tabular In-Context Learner (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.