Bias Detection via Maximum Subgroup Discrepancy

Existing bias detection methods are inefficient: Classical metrics like Total Variation and Wasserstein distances require large amounts of data (high sample complexity) to meaningfully detect bias, making them impractical for real-world applications where data is limited.

A new, more practical approach to bias detection: The paper introduces Maximum Subgroup Discrepancy (MSD), which checks whether biases exist across different groups defined by feature combinations. Importantly, despite the potentially exponential number of possible subgroups, the method only requires data proportional to the number of features—making it computationally feasible.

Bias problems become easier to fix: Unlike "black box" bias metrics, MSD is interpretable, meaning it can clearly identify which specific subgroups are experiencing bias and how large those biases are. This transparency helps practitioners understand and address the root causes rather than just detecting that a problem exists.

Practical implementation is available: The authors provide a concrete algorithm based on Mixed-Integer Optimization (MIO) that organizations can actually use to evaluate their datasets and AI system outputs for bias in real-world scenarios.

Addresses both data quality and system outputs: The framework applies to checking whether training data itself is biased as well as whether an AI system's predictions are fair across different demographic groups, covering multiple critical points in the AI development pipeline.