Pull Request Description

What and why?

Adds multiclass (one-vs-rest) support to four sklearn model-validation tests that were
previously binary-only, so customers validating multiclass classifiers get usable results
instead of a skipped/binary-only test (requested via ZD-717).

Per-class probabilities are read from the underlying estimator's predict_proba — the
VMModel wrapper only exposes the positive-class column, which is binary-only — and the true
labels are one-hot encoded with label_binarize in sklearn's sorted-class order to stay
aligned with the predict_proba columns.

• ROCCurve — for >2 classes, plots one OvR ROC curve per class plus a micro-average, each
  with its own AUC; RawData returns per-class fpr/tpr/auc keyed by class label plus a
  micro entry. Docstring updated for multiclass.
• PrecisionRecallCurve — per-class OvR precision-recall curves plus a micro-average.
• ConfusionMatrix — for >2 classes uses the model's argmax class predictions; the
  threshold parameter is binary-only and is ignored for multiclass.
• PopulationStabilityIndex — computes PSI one-vs-rest, one table/plot per class.

Binary behavior is unchanged in all four. Where a full per-class probability matrix isn't
available (metadata-only models, precomputed single-column probabilities, or a shape that
doesn't match the class count), the multiclass path raises a clear SkipTestError rather than
crashing.

How to test

Automated tests

Run the added/updated unit tests (this repo runs unittest via uv):

make test ONLY="tests.unit_tests.model_validation.sklearn.test_ROCCurve tests.unit_tests.model_validation.sklearn.test_PrecisionRecallCurve"

New multiclass coverage: TestROCCurveMulticlass.test_multiclass_one_vs_rest_traces and
TestPrecisionRecallCurveMulticlass.test_multiclass_one_vs_rest_traces (both train a 3-class
XGBoost model and assert one curve per class + a micro-average trace + baseline, and per-class
AUCs above 0.5). Binary tests in the same files should continue to pass.

Note: the multiclass tests require the xgboost optional extra and are skipped without it.

Manual testing

Run notebooks/code_samples/multiclass_classification_tests.ipynb end-to-end against a
3-class model and confirm:

• ROC and PR plots show one curve per class plus a micro-average curve.
• The confusion matrix renders an N×N grid from argmax predictions.
• PSI renders one table/plot per class.
• Re-running with a binary model produces the original single-curve output.

What needs special review?

• Label/column alignment: label_binarize(y, classes=...) must match predict_proba
  column order (sklearn's sorted class labels). Misalignment would silently mislabel per-class
  curves.
• Skip conditions: the SkipTestError fallbacks should skip only genuinely unsupported
  models (no per-class predict_proba, wrong-shaped probability matrix) and not falsely skip
  supported ones.
• ConfusionMatrix: threshold is now ignored for multiclass — confirm this is the desired
  behavior and is clearly documented.

Dependencies, breaking changes, and deployment notes

None. Library-only change, no migrations, no new env vars, no companion frontend/backend PR.
Binary output is unchanged, so no breaking changes.

Shortcut: SC-17128 ·

Release notes

ROC, Precision-Recall, Confusion Matrix, and Population Stability Index tests now support
multiclass classification models, in addition to binary. For models with more than two
classes, ROC and Precision-Recall curves are drawn one-vs-rest (one curve per class plus a
micro-average), the confusion matrix uses the model's predicted class, and PSI is reported per
class.

Checklist

• What and why
• Screenshots or videos (Frontend) — N/A (no UI change)
• How to test
• What needs special review
• Dependencies, breaking changes, and deployment notes
• Labels applied — enhancement
• PR linked to Shortcut
• Unit tests added (Backend)
• Tested locally — test_ROCCurve + test_PrecisionRecallCurve pass (5 tests, incl. multiclass OvR)
• Documentation updated (if required)
• Environment variable additions/changes documented (if required)