CalibrationEngine<T>
Models & Types · AiDotNet.Evaluation.Engines
Engine for analyzing and improving probability calibration of classifiers.
For Beginners
A well-calibrated classifier produces probabilities that match actual outcomes:
- If model says 70% probability, ~70% of such predictions should be positive
- Many classifiers are NOT well-calibrated (e.g., Random Forest, SVM)
- Calibration methods: Platt scaling (sigmoid), Isotonic regression
How It Works
Why calibration matters:
- Decision making: “Should I act if P > 0.8?” requires calibrated probabilities
- Ranking: AUC doesn’t need calibration, but probability thresholds do
- Ensemble: Combining models requires comparable probability scales
Constructors
| Constructor | Summary |
|---|---|
CalibrationEngine(CalibrationOptions options) |
Initializes the calibration engine. |
Methods
| Method | Summary |
|---|---|
CalibrationResult<T> Analyze(T[] probabilities, T[] actuals, int numBins) |
Analyzes the calibration of a classifier. |
T[] IsotonicCalibration(T[] probabilities, T[] actuals) |
Applies isotonic regression calibration. |
ValueTuple<T[], double, double> PlattScaling(T[] probabilities, T[] actuals) |
Applies Platt scaling (sigmoid calibration) to probabilities. |