Does model multiplicity make fair model search easier?

Determine whether, and under what conditions, the existence of a large number of feasible models that satisfy epsilon-relaxed parity constraints across False Positive Rate, False Negative Rate, and Positive Predictive Value for multiple groups reduces the computational or practical difficulty of discovering at least one such fair model.

Background

The analysis suggests sizable fairness regions under realistic relaxations, implying many candidate models may meet multi-metric fairness. The authors explicitly question whether this multiplicity aids the search process for fair models.

This problem concerns the relationship between the size of the feasible set and practical model discovery, potentially informing algorithm design and search strategies in multi-metric fairness optimization.

References

"Our work leaves open an important next step in ensuring fairness across multiple metrics and for multiple groups: ... Further, does having a large number of feasible models make it easier to find one of those models? ... Unfortunately, these questions are beyond our scope."

— The Possibility of Fairness: Revisiting the Impossibility Theorem in Practice  (2302.06347 - Bell et al., 2023) in Section 6, Conclusions and social impact