Summary

This PR adds McFadden's Pseudo R² Parity as a new fairness metric
to the ClassificationMetric class, designed for binary classification
scenarios where scores are predicted probabilities (0~1) and labels
are binary ground-truth values (0 or 1).

Motivation

Existing fairness metrics in AIF360 do not measure whether a model's
explanatory power is equitable across groups. McFadden's Pseudo R²
provides a log-likelihood-based measure of model fit, and its parity
variant detects disparities between privileged and unprivileged groups.

New Methods

pseudo_r2(self, privileged=None)

$$R^2_{McFadden} = 1 - \frac{\ln L_{model}}{\ln L_{null}}$$

• privileged=True: privileged group
• privileged=False: unprivileged group
• privileged=None: entire dataset

pseudo_r2_parity(self)

$$\Delta R^2 = R^2_{\text{unprivileged}} - R^2_{\text{privileged}}$$

• 0 = parity
• Positive = better fit for unprivileged group
• Negative = better fit for privileged group

Testing

Unit tests added to tests/test_classification_metric.py:

• test_pseudo_r2_overall
• test_pseudo_r2_privileged
• test_pseudo_r2_unprivileged
• test_pseudo_r2_parity