GPyConform’s documentation

GPyConform extends the GPyTorch library by implementing Conformal Prediction (CP) for Gaussian Process Regression (GPR), providing distribution-free, finite-sample valid Prediction Intervals (PIs) under the sole assumption of data exchangeability.

GPyConform supports both the Transductive (Full) CP and Inductive (Split) CP versions of the framework through a unified interface. In both versions it implements a GPR-specific normalized nonconformity measure [1] that leverages the GP predictive variance to construct adaptive symmetric or asymmetric conformal prediction intervals.

Key Features

• Provably Valid Prediction Intervals: Distribution-free, finite-sample coverage guarantees under minimal assumptions (data exchangeability).

• Two CP Framework Versions: - Transductive (Full) CP for Exact GPs: ExactGPCP - Inductive (Split) CP for any GPyTorch regression model: GPRICPWrapper, plus a model-agnostic InductiveConformalRegressor

• Symmetric and Asymmetric PIs in both frameworks.

• Normalized Nonconformity that leverages the GP predictive variance for tighter, adaptive intervals.

• Unified PI Container + Metrics: PredictionIntervals supports retrieving intervals at multiple confidence levels and evaluating empirical coverage error and interval widths.

• Torch-native + GPU-friendly: Works directly with torch.Tensors and can leverage GPU acceleration.

Note

This documentation focuses primarily on the additional functionality provided by GPyConform. For detailed information on the usage and features of the underlying GPyTorch functionality, please refer to the GPyTorch documentation.

Note

• Transductive CP (ExactGPCP) targets ExactGP models with ``GaussianLikelihood`` and relies on an internal patch to GPyTorch’s DefaultPredictionStrategy (applied automatically by default). You can control patching via the GPYCONFORM_AUTOPATCH environment variable, or call gpyconform.apply_patches() manually.

• Inductive CP does not modify model internals and can be used with any GPyTorch regression model (including approximate/deep GPs and different likelihoods). InductiveConformalRegressor can also be used with non-GPyTorch regressors that provide predictive means/variances.

Contents:

• Getting started
  • Installation
  • Tutorials
  • Transductive (Full) GPR Conformal Prediction
  • Inductive (Split) GPR Conformal Prediction
• The GPyConform package
  • Transductive (Full) Conformal Prediction
  • Inductive (Split) Conformal Prediction
  • Conformal Prediction Intervals
• Citing GPyConform
• References