ConformalClassifier<T>
Models & Types · AiDotNet.UncertaintyQuantification.ConformalPrediction
Implements Conformal Prediction for classification tasks.
For Beginners
Conformal Prediction for classification provides prediction sets with guaranteed coverage, not just a single class prediction.
Key differences from regression conformal prediction:
- Instead of an interval, you get a SET of possible classes
- The set is guaranteed to contain the true class with specified probability
Example with 90% confidence:
- Traditional classifier: “This is a cat” (might be wrong)
- Conformal classifier: “This is {cat, dog}” (90% guaranteed to include correct class)
Benefits:
- When the model is uncertain, you get a larger prediction set (e.g., {cat, dog, rabbit})
- When the model is confident, you get a smaller set (e.g., {cat})
- You can defer to a human expert when the prediction set is too large
This is invaluable for:
- Medical diagnosis: Know when to seek specialist opinion
- Autonomous systems: Know when to hand control back to operator
- Quality control: Flag uncertain cases for manual review
Constructors
| Constructor | Summary |
|---|---|
ConformalClassifier(INeuralNetwork<T> model, int numClasses) |
Initializes a new instance of the ConformalClassifier class. |
Methods
| Method | Summary |
|---|---|
HashSet<int> PredictSet(Tensor<T> input, double confidenceLevel) |
Predicts with a conformal prediction set. |
T ComputeAverageSetSize(Matrix<T> testInputs, double confidenceLevel) |
Computes the average size of prediction sets on a test set. |
T EvaluateCoverage(Matrix<T> testInputs, Vector<int> testLabels, double confidenceLevel) |
Evaluates the empirical coverage on a test set. |
void Calibrate(Matrix<T> calibrationInputs, Vector<int> calibrationLabels) |
Calibrates the conformal classifier using a calibration dataset. |