Files
roam/implpaper-tabpfn.org
2026-06-19 18:09:39 +03:00

38 lines
1.7 KiB
Org Mode
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
* impl/paper-tabpfn
[[id:151d5686-6f40-4158-a59a-b0be94cdc969][impl/research]]
*TabPFN: A Transformer That Solves Small Tabular Classification in a Second*
Hollmann et al. — ICLR 2023 — arXiv:2207.01848
** Problem
Small tabular datasets demand expensive hyperparameter search and still often lose to boosted trees. TabPFN asks: can you eliminate tuning entirely while matching AutoML?
** Core Method
A Transformer pre-trained *offline* on millions of synthetic datasets sampled from structural causal models. At inference, the full training set is passed as context — no gradient updates. The model receives =(X_train, y_train, X_test)= as one sequence and outputs predictions in a single forward pass (in-context learning).
** Evaluation
- 18 OpenML-CC18 datasets + 67 small numerical OpenML datasets
- Up to 1,000 training points, 100 features, 10 classes
- Compared against AutoML systems (Auto-sklearn), XGBoost, random forests
** Key Results
- Outperforms boosted trees on small datasets; matches top AutoML
- 230× faster than AutoML baselines; 5,700× with GPU
- No hyperparameter tuning required
** Limitations
- v1: numerical features only, no missing values, max ~1,000 training samples
- TabPFN v2 (2025, arXiv:2502.17361) lifts most constraints — handles larger datasets, mixed types, missing values
** Relevance to ROLL
- [[id:8f59b736-04ea-4d11-9195-30d125a127f8][impl/paper-beyond-rebalancing]] identifies TabPFN as the best-performing classifier on imbalanced tabular data without rebalancing — it is the current bar to beat
- ROLL's niche (optimizing TPR at a specific FPR threshold) is orthogonal: TabPFN uses no custom loss or ROC objective
- If evaluating on small KEEL datasets (≤1,000 samples), TabPFN is the strongest baseline to include