incerto.conformal.conformalized_quantile_regression

incerto.conformal.conformalized_quantile_regression#

incerto.conformal.conformalized_quantile_regression(quantile_model, calib_loader, alpha, q_low=None, q_high=None)[source]#

Conformalized Quantile Regression (CQR) — Romano, Patterson, and Candes, NeurIPS 2019.

Combines quantile regression with split conformal prediction to produce adaptive prediction intervals with finite-sample coverage guarantees.

The method works by: 1. Training a quantile regression model to predict lower and upper quantiles 2. Computing conformity scores on a calibration set as the maximum violation 3. Using these scores to adjust the quantile predictions

Parameters:

  • quantile_model (Module) – A model that outputs (lower_quantile, upper_quantile) predictions. Expected to output shape (batch_size, 2) where: - [:, 0] is the lower quantile prediction - [:, 1] is the upper quantile prediction

  • calib_loader (DataLoader) – DataLoader for calibration data

  • alpha (float) – Miscoverage rate (e.g., 0.1 for 90% coverage)

  • q_low (Optional[float]) – Lower quantile level (default: alpha/2)

  • q_high (Optional[float]) – Upper quantile level (default: 1 - alpha/2)

Returns:

Function mapping inputs to (lower, upper) prediction intervals

Return type:

predictor

Reference:

Romano, Patterson, and Candes. “Conformalized Quantile Regression.” NeurIPS 2019. https://arxiv.org/abs/1905.03222

Example

>>> # Train quantile model (outputs lower and upper quantiles)
>>> quantile_net = QuantileRegressionNet()
>>> # ... train quantile_net ...
>>> predictor = conformalized_quantile_regression(
...     quantile_net, calib_loader, alpha=0.1
... )
>>> lower, upper = predictor(test_x)
Contents