Each dot is an applicant with a true qualification score y (vertical axis) and a group-correlated proxy feature x (horizontal axis — the two clusters are Group A, the majority in the training data, and Group B, the minority).
predicted = 1 if y − w·x + noise > 0
w = 0.9 × (imbalance − 0.5) × 2 × (1 − mitigation)
Disparate Impact = approvalRate(B) / approvalRate(A)
- Training data imbalance — the more training examples come from Group A, the more the model leans on the group-correlated proxy x instead of the true qualification y: weight w grows.
- Fairness mitigation — a reweighting/constraint pass that pulls w back toward zero, restoring a boundary based on merit alone.
- Decision noise — random uncertainty in the model's own confidence, flipping borderline calls either way.
- By the four-fifths rule used in US employment law, a Disparate Impact ratio below 0.80 is treated as evidence of adverse impact.
This mirrors the article's own framing: Bias = (Data Error) + (Design Flaw) + (Human Interpretation), and mitigation works by increasing Fairness = (Data Diversity) × (Algorithmic Transparency) + (Human Oversight).