Classical ML, validation, metrics, interpretation, debugging · insight-0147
Check a reliability diagram before multiplying predicted probabilities by money
- Evidence type
- empirical finding
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- aviation, insurance
Summary: situation, action, reason, boundary
Situation
A boosted tree, random forest or SVM score is about to be used as a probability in an expected-cost or expected-revenue formula.
Action
Plot predicted versus observed frequency on held-out data with bin counts, and calibrate with Platt scaling or isotonic regression on data not used for training when the curve departs from the diagonal.
Reason
Many learners optimize ranking or margins rather than probabilities; calibration maps their scores to observed frequencies.
Boundary
Random forests, boosted trees, SVMs, naive Bayes and neural networks whose outputs drive decisions.
Symptom or risk
Scores that rank well may be pushed toward the middle or the extremes; expected-value decisions become systematically wrong.
Exceptions
Logistic regression is often well calibrated when well specified and not heavily penalized, but still verify.
Trade-offs
Calibration needs held-out data that could otherwise train the model.
Pitfalls
Reading the diagram's sparse extreme bins as precise; show counts per bin.
Lessons and concepts
Claims and their sources
Boosted trees and SVMs tend to produce distorted probabilities that Platt scaling or isotonic regression can correct; random forests also benefit from calibration.
In the scikit-learn calibration comparison, logistic regression is more likely to be well calibrated by itself (canonical logit link), Gaussian naive Bayes pushes probabilities toward 0 and 1, and random forests rarely output probabilities near 0 or 1.