seniorSupervised Learning
What is calibration vs discrimination in classification models?
Updated May 17, 2026
Short answer
Calibration measures probability accuracy, while discrimination measures class separation ability.
Deep explanation
A model can discriminate well (separate classes accurately) but still be poorly calibrated (incorrect probabilities). Discrimination is measured using ROC-AUC, while calibration is measured using Brier score or calibration curves. Both are important in risk-sensitive applications like healthcare.
Real-world example
A medical model correctly ranking patients by risk but overestimating probabilities.
Common mistakes
- Assuming high ROC-AUC means good probability estimates.
Follow-up questions
- Can a model have high AUC but poor calibration?
- How do you fix calibration issues?