Apply regression target transforms to the unnormalized target [RES-2639]

[bejaeger]

Closes RES-2639. Stacked on #1198 (targets that branch, so this diff is only the behaviour change); review #1198 first.

Change

One swap, in make_target_transform: the preset reshapes the target in its original units, and the result is standardized afterwards.

-            (STANDARDIZE_STEP, StandardizeTarget()),
             (TARGET_TRANSFORM_STEP, transform),
+            (STANDARDIZE_STEP, StandardizeTarget()),

Applied to a z-normalised target, 1_plus_log was log1p of a z-score — undefined below -1, so for a symmetric target ~15% of rows came out non-finite and reached the model, the log presets carrying no scaler of their own. safepower was fitted on standardized values, where yeo-johnson cannot reach the log limit that removes skew from a heavy right tail.

Also resolves the "none" preset to no transform (it is the identity, so composing it only costs a pass over the target), and fixes the _transform_labels_one docstring, which called its input "unprocessed" — the ticket's second item.

Impact

Changes the v2 / v2.5 / v3 defaults, whose ensembles use safepower for half of their estimators. Unaffected: the v2.6 default ("none"), n_estimators=1 on any version, and every estimator whose target transform is None.

Validation

TabArena regression, 13 datasets, 222 splits per arm, 8 estimators, checkpoint x49jl1p6, 1332 fits, no failures (full report on the ticket):

comparison	rmse	mae	1−r2	better (rmse)
this change vs base	−0.031%	−0.069%	−0.062%	9/13
no power transform vs base	+0.069%	−0.080%	+0.140%	5/13

Errors, so negative is better. Gains concentrate on heavy-tailed positive targets: healthcare_insurance_expenses −0.23%, diamonds −0.16%, physiochemical_protein −0.13%.

That run measured these semantics through an earlier implementation (80bf5fa0, which composed the frame into each pipeline instead of moving the standardization). This stack reproduces it to 1e-7 relative on the five datasets used to cross-check, so the validation transfers.

Caveat from that report, carried over: dropping the power transform entirely costs only +0.069% rmse on those 13 datasets, so that view caps what any change to how the transform composes can be worth, and this effect sits below its noise floor — for scale, incidental numerical drift between v8.3.0 and base moved one dataset by 0.60%. The direction is right and the mechanism is a bug fix; the magnitude is not certifiable on TabArena regression. Suggested follow-up on the ticket: validate on skewed positive-target data.

A variant that additionally scale-normalises the yeo-johnson transforms was measured in the same run and is not part of this stack — it gave back about half the gain (+0.017% rmse). Parked on benjamin/res-2639-safepower-scale-normalized.

Tests

The test that #1198 pinned to the old ordering is flipped here, which is the reviewable core of the diff. Added: log1p of the target is finite where log1p of a z-score is not, the "none" preset resolves to no transform, and end-to-end accuracy on a skewed target for three presets.

🤖 Generated with Claude Code

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit e7d29d7. Configure here.}

[cursor]

Comment narrates prior transform order

Low Severity

The comment on the new pipeline order explains what the previous ordering did differently (1_plus_log as log1p of a z-score). Team convention is that comments describe only current behavior, not prior behaviour or ticket history.

 

^{Triggered by learned rule: Docstrings describe current behavior only — no call sites or code history}

^{Reviewed by Cursor Bugbot for commit e7d29d7. Configure here.}