fix: handle both 1D and 2D input shapes in XGBoost custom objectives

[Soutehkeshan]

Summary

Fixes compatibility with XGBoost 3.2, which changed the shape of arrays passed to custom objectives via the sklearn interface. With multi_strategy="one_output_per_tree", y_true and y_pred are now passed as 2D (n_samples, n_quantiles) arrays instead of 1D flattened arrays of length n_samples * n_quantiles.

Closes #865

Root Cause

The old reshape logic in both objective functions assumed 1D input:

n_items = len(y_true)
n_quantiles = len(quantiles)
n_rows = n_items // n_quantiles  # ← breaks with 2D input
y_pred = np.reshape(y_pred, (n_rows, n_quantiles))
y_true = np.reshape(y_true, (n_rows, n_quantiles))
sample_weight = np.reshape(sample_weight, (n_rows, -1)) if sample_weight is not None else None

With 2D input, len(y_true) returns n_samples (not n_samples * n_quantiles), so n_rows = n_samples // n_quantiles is wrong, causing the reshape to fail or produce incorrect results.

The below image shows running a minimal script using XGBoost 3.1:

The below image shows running the same minimal script using XGBoost 3.2:

Fix

Replaced with shape-agnostic normalisation in arctan_loss_multi_objective and pinball_loss_multi_objective:

n_quantiles = len(quantiles)
y_pred = np.reshape(y_pred, (-1, n_quantiles))
y_true = np.reshape(y_true, (-1, n_quantiles))
sample_weight = sample_weight[:, np.newaxis] if sample_weight is not None else None

np.reshape(arr, (-1, n_quantiles)) is a no-op when the array is already (n_samples, n_quantiles) and correctly reshapes a 1D flat array, maintaining backward compatibility with XGBoost <3.2.

Investigation Notes

• mean_pinball_loss: already used np.reshape(y_pred, [-1, n_quantiles]) which handles both shapes correctly — no change needed
• metrics_probabilistic.py: no changes needed, sample_weight_eval_set verified working correctly with XGBoost 3.2
• xgboost_forecaster.py: no changes needed
• pyproject.toml: already at >=3,<4 on this branch — no change needed

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[egordm]

LGTM, both affected objective functions are correctly fixed and the pattern is consistent. Approving.

[egordm]

Hi @Soutehkeshan, thanks so much for picking this up and submitting a fix so quickly! Really appreciated.

The code change looks correct. Both arctan_loss_multi_objective and pinball_loss_multi_objective are now handled, and the existing tests pass. The only thing missing is test coverage for the new 2D input shape so we have a regression guard going forward. To avoid blocking your PR, I've added those tests in a separate follow-up (PR #867) that we'll merge right after this one.

On the DCO check failing: if you haven't already, you can either add a sign-off by amending your commit (git commit --amend --signoff) or we can handle it with a squash-merge on our side. No worries either way, we'll sort it out.

Looking forward to seeing more contributions from you!

[Soutehkeshan]

Hi @egordm,

Thanks for your feedback. I'm not sure if sign-off is still needed considering that the PR has already been merged. If so, could you please handle it with a squash-merge on your side?

Looking forward to contributing more.