Bias mitigation reduces systematic distortion or unfairness in model outputs. It treats models as tools whose results can become “crooked” when developers or users overlook bias, while recognizing that careless correction can overcompensate and make results unusable. section 8.5
Use an output-focused process to identify and address possible bias:
- Monitor outputs regularly for trends and patterns. section 8.5.2
- Compare performance across groups or data subsets, and seek feedback from users or stakeholders. section 8.5.2
- Apply disparate impact, confusion matrix, or residual analysis where appropriate. section 8.5.2
- Consider post-processing, including reranking, resampling, or reweighing, without treating it as a substitute for reassessing the model’s suitability. section 8.5.2
Residential loan underwriting
A model provides a residential-loan applicant’s probability of default for an underwriter’s final lending decision. Its outputs should be evaluated for bias arising from skewed historical training data before relying on them. section 8.5.2
