Bias Mitigation
Once bias is detected, FairnessAudit offers automated remediation pipelines using state-of-the-art algorithms to correct the model's behavior while attempting to preserve overall accuracy.
Correlation Remover
A pre-processing technique that projects features into a subspace orthogonal to the sensitive attributes. This removes linear correlations between features and protected classes before training even begins.
Threshold Optimizer
A post-processing technique that adjusts the decision boundaries (thresholds) for different sensitive groups in a trained model to enforce Demographic Parity or Equalized Odds.
Accuracy vs. Fairness Trade-off
Applying mitigation techniques usually results in a slight drop in overall model accuracy. FairnessAudit automatically computes the "Cost of Fairness" by displaying a side-by-side comparison of accuracy metrics before and after mitigation, allowing you to make an informed business decision.