[SC-17187] ZD-717: GINITable fails on multiclass models (follow-up to PR #535)

[juanmleng]

Pull Request Description

What and why?

GINITable (validmind/tests/model_validation/statsmodels/GINITable.py) computed AUC/GINI/KS with binary-only sklearn calls (roc_curve, roc_auc_score) against a single positive-class probability column. On a model with more than two classes it raised ValueError: multiclass format is not supported.

This adds a one-vs-rest path for multiclass models, mirroring the approach PR #535 used for the sklearn tests (ROCCurve, PrecisionRecallCurve, ConfusionMatrix, PopulationStabilityIndex). #535 fixed those four but did not touch statsmodels.GINITable; this is the remaining follow-up.

After this change:

• Binary models: unchanged — one row of AUC/GINI/KS.
• Multiclass models: the metrics are computed one-vs-rest and the table returns one row per class plus a micro-average row (with a Class column). Per-class probabilities are read from the underlying estimator's predict_proba, since the VMModel wrapper exposes only the positive-class column.
• Models that can't produce a full per-class probability matrix (metadata-only models, or predictions supplied as a single precomputed probability column) raise SkipTestError instead of crashing.

How to test

Run the existing unit tests:

python -m pytest tests/unit_tests/model_validation/statsmodels/test_GINITable.py

Or reproduce the original failure and confirm the fix directly (this is the snippet from the repro notebook used to validate the change):

import numpy as np
import pandas as pd
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression
import validmind as vm
from validmind.tests.model_validation.statsmodels.GINITable import GINITable


def build(n_classes, seed=42):
    X, y = make_classification(
        n_samples=300, n_features=5, n_informative=4, n_redundant=0,
        n_classes=n_classes, n_clusters_per_class=1, random_state=seed,
    )
    df = pd.DataFrame(X, columns=[f"f{i}" for i in range(5)])
    df["target"] = y
    model = LogisticRegression(max_iter=1000).fit(X, y)
    ds = vm.init_dataset(input_id=f"ds{n_classes}", dataset=df, target_column="target", __log=False)
    m = vm.init_model(input_id=f"m{n_classes}", model=model, __log=False)
    ds.assign_predictions(m)
    return ds, m


# Binary — worked before, still works
ds2, m2 = build(2)
print(GINITable(ds2, m2)[0])

# Multiclass (3 classes) — raised "ValueError: multiclass format is not supported"
# before this PR; now returns one row per class plus a micro-average
ds3, m3 = build(3)
print(GINITable(ds3, m3)[0])

Expected multiclass output:

   Class       AUC      GINI        KS
0      0  0.958650  0.917300  0.845000
1      1  0.926300  0.852600  0.725000
2      2  0.933950  0.867900  0.785000
3  micro  0.940167  0.880333  0.768333

What needs special review?

The multiclass path reads probabilities from the underlying estimator (model.model.predict_proba) rather than the VMModel wrapper's y_prob, because the wrapper returns only the positive-class column. Worth confirming this is the right access pattern (it matches ROCCurve from #535) and that the SkipTestError fallback covers the model types you expect (metadata-only / precomputed single-column probabilities).

Dependencies, breaking changes, and deployment notes

Follow-up to #535. No breaking changes — binary output is unchanged.

Release notes

GINITable now supports multiclass classification models. For models with more than two classes it reports AUC, GINI, and KS one-vs-rest (one row per class plus a micro-average) instead of failing with multiclass format is not supported. Binary models are unaffected.

Checklist

• What and why
• Screenshots or videos (Frontend)
• How to test
• What needs special review
• Dependencies, breaking changes, and deployment notes
• Labels applied
• PR linked to Shortcut
• Unit tests added (Backend)
• Tested locally
• Documentation updated (if required)
• Environment variable additions/changes documented (if required)

[github-actions]

Pull requests must include at least one of the required labels: internal (no release notes required), highlight, enhancement, bug, deprecation, documentation. Except for internal, pull requests must also include a description in the release notes section.

[github-actions]

PR Summary

This PR extends the functionality of the GINITable module by adding support for multiclass classification metrics. The key changes include:

• Introducing a new internal function _multiclass_gini_table which computes AUC, GINI, and KS metrics for multiclass models using a one-vs-rest approach. The function obtains the full per-class probability matrix from the underlying model (via predict_proba) and verifies that the returned shape matches the number of classes.
• Updating the main GINITable function to delegate to the new multiclass logic when the dataset contains more than two unique classes.
• Implementing robust error handling: if the underlying model does not support predict_proba or if the output shape is not as expected, a SkipTestError is raised to gracefully bypass unsupported scenarios.
• Adding a comprehensive set of unit tests within the TestGINITableMulticlass class. These tests cover both successful metric computation (checking that per-class and micro-average metrics are correctly calculated and meet expected relationships) and the proper triggering of the SkipTestError when a model cannot produce per-class probabilities.

Overall, these changes improve the library's capability to handle multiclass model evaluations while maintaining backward compatibility with binary classification scenarios.

Test Suggestions

• Test models with binary classification to ensure existing functionality remains unchanged.
• Simulate a multiclass scenario where the model's predict_proba returns an improperly shaped array to trigger the SkipTestError.
• Test models with a predict_proba method that raises an exception to verify that the error handling in _multiclass_gini_table works as expected.
• Validate that the per-class and micro-average metrics, including AUC, GINI, and KS, are computed correctly for diverse datasets.