SplitConformalPredictor<T>
Models & Types · AiDotNet.UncertaintyQuantification.ConformalPrediction
Implements Split Conformal Prediction for regression tasks.
For Beginners
Conformal Prediction is a framework that provides prediction intervals with guaranteed coverage, regardless of the underlying model.
Key concepts:
- Instead of a single prediction, you get a prediction interval: [lower_bound, upper_bound]
- The interval is guaranteed to contain the true value with a specified probability (e.g., 90%)
- This guarantee holds for ANY model (neural network, random forest, etc.)
Example: If you set confidence level = 90%:
- The model predicts: “House price will be between $180K and $220K”
- You’re guaranteed that at least 90% of such intervals contain the true price
How it works:
- Split data into training set, calibration set, and test set
- Train your model on the training set
- Use the calibration set to compute “non-conformity scores” (prediction errors)
- Use these scores to build prediction intervals for new data
This is particularly valuable when you NEED reliability guarantees, such as:
- Medical diagnosis (must know uncertainty bounds)
- Safety-critical systems (must guarantee coverage)
- Scientific applications (need statistically valid uncertainty)
Constructors
| Constructor | Summary |
|---|---|
SplitConformalPredictor(IModel<Tensor<T>, Tensor<T>, ModelMetadata<T>> model) |
Initializes a new instance of the SplitConformalPredictor class. |
Methods
| Method | Summary |
|---|---|
T ComputeAverageIntervalWidth(Matrix<T> testInputs, double confidenceLevel) |
Computes the average interval width on a test set. |
T EvaluateCoverage(Matrix<T> testInputs, Vector<T> testTargets, double confidenceLevel) |
Evaluates the empirical coverage of prediction intervals on a test set. |
ValueTuple<T, T, T> PredictWithInterval(Tensor<T> input, double confidenceLevel) |
Predicts with a conformal prediction interval. |
void Calibrate(Matrix<T> calibrationInputs, Vector<T> calibrationTargets) |
Calibrates the conformal predictor using a calibration dataset. |