Research map / Label-efficient and robust learning / Test-time adaptation
TabLog: Test-Time Adaptation for Tabular Data Using Logic Rules
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 TabLog, a test-time adaptation method for tabular classifiers that keeps the logical structure of a learned rule ensemble fixed and adapts only discretization thresholds and rule weights using a binning-based contrastive loss on unlabeled target data.
Problem
Test-time (source-free) adaptation of predictive models trained on heterogeneous tabular data under distribution shift, where existing TTA methods struggle with mixed feature types and complex feature dependencies.
Approach
Learn a logical neural network rule ensemble with learnable discretization thresholds on the source domain; at test time freeze the rule structure and adapt only thresholds and rule voting weights by minimizing entropy plus a binning-informed InfoNCE contrastive loss on unlabeled target data.
Key results
- On TableShift ASSISTments, TabLog reaches 62.64% accuracy and 60.96% macro-F1, above the best baseline TAST (61.01% / 58.44%).
- On Sepsis, TabLog reaches 98.78% accuracy versus 92.46% for TAST and XGBoost.
- TabLog outperforms all listed baselines on the four TableShift datasets and is best or comparable on the simulated-shift datasets (e.g., Airbnb 66.49% vs TAST 65.52% accuracy).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TabLog’s rule-ensemble test-time adaptation instead of FT-trans TAST | Tabular test-time adaptation · ASSISTments | accuracy: 61.01±0.31 -> 62.64±0.54 | Helped |
| Use TabLog’s rule-ensemble test-time adaptation instead of FT-trans TAST | Tabular test-time adaptation · Sepsis | accuracy: 92.46±1.08 -> 98.78±0.19 | Helped |
| Use TabLog’s rule-ensemble test-time adaptation instead of FT-trans TAST | Tabular test-time adaptation · Hospital Readmission | Macro-F1: 61.73±0.28 -> 62.81±0.69 | Helped |
| Use TabLog’s rule-ensemble test-time adaptation instead of FT-trans TAST | Tabular test-time adaptation · PhysioNet | accuracy: 88.36±0.79 -> 89.54±0.47 | Helped |
| Use TabLog’s rule-ensemble test-time adaptation instead of FT-trans TAST | Tabular test-time adaptation · Airbnb | accuracy: 65.52±8.01 -> 66.49±0.92 | Helped |
| Use TabLog’s rule-ensemble test-time adaptation instead of FT-trans TAST | Tabular test-time adaptation · Channel | accuracy: 36.37±0.60 -> 37.96±0.51 | Helped |
| Use TabLog’s rule-ensemble test-time adaptation instead of FT-trans TAST | Tabular test-time adaptation · Jigsaw | accuracy: 95.91±0.00 -> 95.17±0.82 | Hurt |
| Use TabLog’s rule-ensemble test-time adaptation instead of XGboost | Tabular test-time adaptation · Wine | accuracy: 16.92±0.35 -> 15.98±0.47 | Hurt |
| Increase the number of output nodes (rule count proxy) (ablation) | TabLog · ASSISTments | Macro-F1 | Mixed |
| Vary the number of bins for contrastive sample generation (ablation) | TabLog · ASSISTments | Macro-F1 | Mixed |
| Set λ near 0.1 for the self-supervised loss (ablation) | TabLog · Natural-shift benchmark datasets | Macro-F1 | Helped |
Surprising findings
- Tent, EATA and SHOT with an MLP backbone often underperform or are close to chance on tabular data, while the FT-Transformer backbone helps them substantially.
Limitations
- Performance is sensitive to the number of rules and bins, with optimal settings being data dependent.
- Assumes only marginal covariate shift; does not handle label shift or shifts in feature combinations.
- Small datasets remain challenging for all methods, including TabLog.
Experimental setup
Datasets: ASSISTments Sepsis Hospital Readmission PhysioNet Airbnb Channel Jigsaw Wine
Benchmarks: ASSISTments Sepsis Hospital Readmission PhysioNet Airbnb Channel Jigsaw Wine
Metrics: accuracy macro-F1
Baselines: Logistic Regression XGBoost CatBoost VIME TransTab Scarf Tent EATA SHOT TAST
Builds on: Test-time training (SHOT) Logical neural networks Scarf TAST Tent
Closest papers by meaning
- Fully Test-time Adaptation for Tabular Data (AAAI 2025 · similarity 0.91)
- TabFSBench: Tabular Benchmark for Feature Shifts in Open Environments (ICML 2025 · similarity 0.86)
- TabDPT: Scaling Tabular Foundation Models on Real Data (NeurIPS 2025 · similarity 0.86)
- A Data-Centric Perspective on Evaluating Machine Learning Models for Tabular Data (NeurIPS 2024 · similarity 0.86)
- Drift-Resilient TabPFN: In-Context Learning Temporal Distribution Shifts on Tabular Data (NeurIPS 2024 · similarity 0.86)
- TabPrep: Closing the Feature Engineering Gap in Tabular Benchmarks (NeurIPS 2026 · similarity 0.86)
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.