feat: add predict_uncertainty() for ensemble-based regression intervals

[Debi-Prasad-Panda]

Summary

Adds automl.predict_uncertainty(X, alpha=0.05) for regression tasks, as discussed in the issue thread.

This method computes ensemble-based uncertainty intervals using predictions from all sub-models selected in the Ensemble. For each sample:

• prediction = weighted mean of sub-model predictions
• prediction_variance = weighted variance of sub-model predictions
• prediction_std = sqrt(prediction_variance)
• lower = prediction - z * prediction_std
• upper = prediction + z * prediction_std

where z = norm.ppf(1 - alpha/2) (e.g., z ≈ 1.96 for alpha=0.05).

Important: This is documented as an ensemble disagreement interval, not a statistically calibrated confidence interval.

Changes

• supervised/automl.py — Added public predict_uncertainty() method with full docstring.
• supervised/base_automl.py — Added internal _predict_uncertainty() with:
  • Regression-only validation
  • Automatic Ensemble fallback when best model is a single algorithm
  • Weighted mean/variance/std computation across sub-models
  • Z-score based interval bounds
• tests/tests_automl/test_predict_uncertainty.py — 5 unit tests covering:
  • Before fit
  • Classification task (rejected)
  • No Ensemble available (rejected)
  • Ensemble fallback from single best model
  • Full regression success with output validation

Output Format

prediction	prediction_std	prediction_variance	lower	upper

[Debi-Prasad-Panda]

Hi @pplonski, apologies for the delay on this — I had a family loss. I've now opened a PR implementing predict_uncertainty() as discussed.

It follows the exact spec you outlined — weighted mean, variance, std from ensemble sub-model predictions, with z-score based intervals from the alpha parameter. The output is a DataFrame with columns: prediction, prediction_std, prediction_variance, lower, upper.

I've also added an automatic fallback so it works even when the best model selected by AutoML is a single algorithm (it finds and uses the trained Ensemble internally).

5 unit tests are included covering all edge cases. Looking forward to your feedback!