Proxy Logic
Simply removing a sensitive attribute (like race or gender) from your dataset does not guarantee fairness. Models can easily infer this information from other variables, known as proxy features.
Redlining & Hidden Bias
A classic example of proxy bias is "redlining" in banking. A bank might not use race to determine loan eligibility, but they use ZIP codes. Because ZIP codes in many regions are highly correlated with race, the ZIP code becomes a proxy for race, leading to identical discriminatory outcomes.
Automated Detection
FairnessAudit automatically computes mutual information and Pearson correlation coefficients between all features and the defined sensitive attribute to detect hidden proxies.
Warning Thresholds
Features that show a correlation higher than 0.7 with the sensitive attribute are immediately flagged in the UI, allowing engineers to remove or transform them before training.