Glossary
Updated
The use of quantitative and qualitative techniques (statistical disparity tests, counterfactual simulations, subgroup performance comparisons, and error-analysis dashboards) to reveal where and how models treat different cohorts unequally. Bias detection is continuous: as data evolves, new biases may emerge, requiring re-evaluation at regular intervals.
Example
An e-commerce company runs its product-recommendation model through a bias-detection pipeline every quarter, checking if certain customer demographics receive fewer or lower-quality suggestions. When the Hispanic segment’s click-through rate lags behind others, data scientists retrain the model with balanced user-behavior samples to correct the disparity.