A proposal of smooth interpolation to optimal transport for restoring biased data for algorithmic fairness

Biased training data, not just flawed algorithms, is a major source of algorithmic bias: The paper emphasizes that AI systems often produce unfair results because they're trained on data that reflects existing societal biases—meaning the problem originates upstream from the algorithm itself and must be addressed at the data level.

A mathematical technique called optimal transport can help "repair" biased datasets: The researchers propose using optimal transport theory to adjust feature distributions between privileged and unprivileged groups, making the underlying data more similar across demographic groups so algorithms trained on it produce fairer outcomes.

The proposed "Extended Total Repair" (ExTR) method makes bias correction more practical for real-world use: Rather than recalculating fairness corrections every time new data arrives (which is computationally expensive), the ExTR method uses smooth interpolation to efficiently update bias corrections incrementally, making the approach scalable for organizations continuously receiving new data.

The approach focuses on two-group fairness scenarios: The method specifically addresses cases where data divides into "privileged" and "unprivileged" groups (such as different genders or ethnic backgrounds), working to ensure that other attributes don't correlate with these sensitive demographic characteristics.