Extend Empirical Differential Fairness metric

[jetverbeek]

The current implementation of the EDF metric only returns a single scalar value summarizing fairness across protected attributes. For intersectional analysis it would be highly valuable to identify which combinations of attribute groups contribute most to the maximum log-ratio.

Extend the metric so that it returns:

• The pair(s) of attribute groups that produce the max log ratio.
• A breakdown of all/top X group combinations and their respective log-ratios.

[Hanabi9248]

Could you assign this to me? I would keep smoothed_empirical_differential_fairness() returning a scalar and add a separate method for the group-pair breakdown, including the pairs attaining the maximum and an optional top-k limit. The calculations would use the same instance weights and Dirichlet smoothing as the existing metric, with tests for ties and multiple protected attributes. Please let me know if you prefer an optional details argument on the existing method instead.