Model Evaluation, Validation, Bias & Ethics

Assessing Models: Validation, Bias and Ethics

Correct evaluation determines whether a model actually works, while ethics determine if it can be applied without harm or discrimination.

Validation protocols include cross-validation, stratification, and time splits. Quality metrics are chosen based on the task and business goals; there's a trade-off between precision and recall.

Calibration of Probabilities: Platt Scaling, Isotonic Regression

The ‘accuracy’ metric can be misleading on imbalanced classes. Incorrect splits lead to inflated quality assessments. There's a lack of a formal process for challenging decisions in high-risk domains.

Evaluation isn't just one number; it’s a set of practices that ensure the model is honest, correct, and legitimate.

High Test Metrics Don’t Guarantee Model Responsibility

1) Choose appropriate validation and stratification protocols. 2) Measure multiple metrics with different trade-offs. 3) Verify fairness for subgroups. 4) Document limitations and risks, implement a challenge process.

In credit scoring, in addition to AUC, the difference in rejection rates between demographic groups is measured, probabilities are calibrated, and decisions are made considering ethical policies and regulatory requirements.

Frequently asked questions