Overview
Machine learning models are frequently deployed in environments where data characteristics diverge from their calibration datasets. Conformal prediction and its risk-control generalizations offer distribution-free guarantees during deployment. However, these guarantees typically rely on assumptions of exchangeability or precise knowledge of the likelihood ratio in cases of covariate shift. The practical scenario, where the likelihood ratio must be estimated, has been an open problem, particularly for Bayesian constructions available under exchangeability.
This research addresses this gap by developing Weighted Bayesian Conformal Prediction. This method pushes a Dirichlet process prior on the calibration distribution through an estimated likelihood ratio to achieve an exact Bayesian posterior over the deployed risk. While the posterior is exact given a weight function, this function itself is estimated during deployment. Consequently, the study also establishes finite-sample upper and lower bounds on the risk incurred by the selected threshold, considering both weight-estimation error and conditional shift.
Research Context
The challenge in deploying machine learning predictors often stems from the mismatch between training and deployment data distributions. Conformal prediction (CP) offers a framework to quantify uncertainty and provide valid prediction sets with distribution-free coverage guarantees. Its generalizations extend these risk-control capabilities. A critical limitation, however, arises when these guarantees are contingent on data exchangeability or a precisely known likelihood ratio in covariate shift scenarios.
The specific open problem targeted by this work is the case where the likelihood ratio in a covariate shift needs to be estimated in practice. Previous Bayesian constructions were not applicable under these conditions. The research aims to provide a robust Bayesian framework that can operate effectively even when the shift parameters are not perfectly known.
Approach
The core of the methodology involves starting from a Dirichlet process prior applied to the calibration distribution. This prior is then propagated through the likelihood ratio to compute the exact Bayesian posterior over the deployed risk. This construction is designed to extend Bayesian guarantees to practical scenarios involving covariate shift where the likelihood ratio is not known a priori but must be estimated.
Acknowledging that the weight function, which encapsulates the likelihood ratio, is estimated at deployment, the research further provides theoretical guarantees. Specifically, finite-sample upper and lower bounds are proven for the actual risk incurred by the chosen prediction threshold. These bounds are expressed in terms of the error associated with weight estimation and the conditional shift observed in the data.
Findings
- The proposed method, Weighted Bayesian Conformal Prediction, yields an exact Bayesian posterior over the deployed risk, even when the likelihood ratio for covariate shift is estimated. This is achieved by pushing a Dirichlet process prior through the estimated likelihood ratio.
- Finite-sample upper and lower bounds on the actual risk incurred by the selected threshold were proven. These bounds directly relate to the error in estimating the weight function and the nature of the conditional shift.
- On both synthetic and real covariate shifts, across regression and classification tasks, the realized failure rate using this method remained close to the target failure rate. This outcome contrasts with scenarios where shift-blind selection methods failed to maintain target rates.
- The output of the deployment process is a posterior distribution over the risk of every possible threshold, rather than a single numerical risk estimate.
Why This Matters
The ability to provide distribution-free guarantees for machine learning predictors under estimated covariate shifts addresses a significant practical challenge in model deployment. By ensuring the realized failure rate aligns with the target rate, even when data distributions change, the method offers increased reliability and interpretability for deployed models. The provision of a posterior over risk for all thresholds, rather than a point estimate, allows for a more nuanced understanding of prediction uncertainty, which is crucial for decision-making in critical applications.