H2ML¶
A 4-step AutoML pipeline for tabular data that wraps sklearn-compatible estimators. Given a feature matrix and target, it screens all registered models, reduces features via SHAP importance and correlation filtering, and tunes the winner with Optuna — all in one call.
How it works¶
| Step | What happens |
|---|---|
| 1 | K-fold CV all models (× optional y-transforms) on all features |
| 2 | SHAP importance → correlation-based feature drop |
| 3 | K-fold CV all models on reduced features; compare vs step 1 |
| 4 | Optuna HPO on the winning (model, stage, transform) |
Features¶
- Model screening — all sklearn classifiers and regressors in one parallel CV pass
- Optional boosting — LightGBM, XGBoost, CatBoost registered when installed
- SHAP feature selection — out-of-fold SHAP avoids leaking importance scores
- Spatial CV — block or AHC-based splitters for geospatial data
- y-transform sweep — log, sqrt, and winsorized variants evaluated jointly
- Conformal prediction — guaranteed-coverage intervals built from OOF residuals
- Persistence — Parquet + npy + joblib; reload and resume any pipeline stage