Uncovering Discrimination Clusters: Quantifying and Explaining Systematic Fairness Violations

Individual fairness checks miss systemic patterns: The paper reveals that traditional fairness audits focusing on isolated cases of discrimination (comparing two similar individuals with different outcomes based on protected attributes) fail to detect broader patterns where entire groups experience arbitrary treatment based on race, gender, or other protected characteristics.

"Discrimination clustering" identifies hidden bias patterns: The research introduces a new concept that goes beyond pairwise comparisons to find regions in datasets where small changes in protected attributes lead to dramatically different outcomes for groups of people, exposing systematic bias that standard fairness metrics overlook.

Formal verification combined with search is more effective: HyFair, the technique presented in the paper, combines mathematical verification methods with randomized search to both certify when systems meet fairness standards and actively hunt for discriminatory clusters, providing both guarantees and practical discovery of bias.

Algorithmic arbitrariness based on protected attributes is the core problem: The paper emphasizes that the real issue isn't just unequal outcomes, but unexplainable variation in how algorithms treat people who differ only in protected attributes—suggesting that even if outcomes seem reasonable individually, the pattern of decisions reveals unfair arbitrariness.