Y-Transforms¶

Y-transforms allow the pipeline to evaluate whether predicting in a transformed target space improves generalisation. Steps 1 and 3 sweep all (model × transform) combinations; the winning pair is selected once and used throughout.

Usage¶

result = pipeline.run(store, transforms=["log", "sqrt", "count", "winlog"])
print(result.y_transform)   # e.g. "log"

Only supported for regression (task_type="regression").

Available transforms¶

Name	Transform	Inverse	Notes
`"count"`	identity	identity	Baseline — no transformation
`"log"`	`log1p(y)`	`expm1(y)`	Requires non-negative y
`"sqrt"`	`sqrt(y)`	`y²`	Requires non-negative y
`"wincount"`	winsorize(y)	identity	Clips upper outliers to IQR upper limit
`"winlog"`	winsorize(log1p(y))	`expm1(y)`	Winsorize then log
`"winsqrt"`	winsorize(sqrt(y))	`y²`	Winsorize then sqrt

Winsorize-based transforms return None when no upper outliers are found and are silently skipped by build_transform_stores. If all winsorized variants are skipped, only non-winsorized transforms are evaluated.

Metrics in transformed space¶

CV metrics in steps 1 and 3 are computed in the original scale when y_true is provided — the inverse transform is applied to predictions before metric computation. This ensures R2, MAE, and RMSE are comparable across transforms.

Conformal intervals and transforms¶

When result.y_transform is set, FinalModel.predict_interval() returns bounds in the transformed space. Apply the inverse to get original-scale intervals:

from h2ml.preprocessing.transforms import INVERSE_TRANSFORMS

lower, upper = final.predict_interval(X_new, alpha=0.10)
inv = INVERSE_TRANSFORMS[result.y_transform]
lower_orig, upper_orig = inv(lower), inv(upper)