JAWS: Auditing Predictive Uncertainty Under Covariate Shift

Andrew Prinster, Anqi Liu, Suchi Saria

Introduction

Most machine learning models rely on the assumption of exchangeability—the expectation that future, unseen data will resemble the training data. In real-world applications, this assumption frequently collapses due to covariate shift, where input features (such as patient demographics or environmental conditions) change over time.

When distribution shift occurs, standard uncertainty quantification methods like conformal prediction falter. For instance, methods designed to provide a 90% confidence interval that actually contains the true value 90% of the time can see their coverage drop significantly below the target threshold when faced with biased data. To address this, researchers have introduced JAWS (JAckknife+ Weighted), a suite of wrapper methods that provide reliable uncertainty estimates and audit model reliability even under covariate shift.

The JAWS Methodology

The JAWS framework introduces several key components to restore theoretical guarantees and improve computational efficiency:

  1. JAW (JAckknife+ Weighted): The core method adapts the standard jackknife+ (a leave-one-out resampling technique) by incorporating data-dependent likelihood-ratio weights. These weights effectively "re-balance" the training data to better match the test distribution, restoring finite-sample coverage guarantees even under shift.
  2. JAWA (JAW Approximation): Standard jackknife methods are computationally expensive because they require retraining the model $n$ times (once for every data point). JAWA utilizes higher-order influence functions—akin to a sophisticated Taylor series approximation—to estimate the retrained model's parameters without actual retraining. This makes the method orders of magnitude faster while maintaining asymptotic guarantees.
  3. Error Assessment (JAW-E & JAWA-E): While traditional tools output a predictive interval ("the answer is between X and Y"), the authors propose a "reverse" task. Given a user-defined safety threshold (e.g., "the prediction must be within 10% of the true value"), these methods estimate the probability that a prediction is safe.
Summary of the JAWS methods, their tasks, and theoretical guarantees under covariate shift.

Key Insights and Experimental Results

The researchers evaluated JAWS on five real-world UCI regression datasets (including Airfoil self-noise, Red wine quality, and Superconductivity), inducing covariate shift via exponential tilting to simulate real-world bias.

Weighting Restores Theory: A major insight of the paper is that importance weighting can be seamlessly integrated into the jackknife+ framework. This bridges classical importance sampling with modern conformal inference, maintaining coverage guarantees without producing overly conservative, massive intervals. In experiments, while standard baselines saw coverage drop to 75-85% under shift, JAW consistently maintained the target 90% coverage with significantly narrower intervals.

Approximation is Highly Effective: The empirical results demonstrate that full model retraining is unnecessary. Using influence functions, JAWA achieves performance practically indistinguishable from the full JAW method but at a fraction of the computational cost, making it viable for large-scale deep learning models.

The Power of the "Reverse" Task: In safety-critical domains, knowing the probability that a prediction falls outside a safe tolerance zone is often more actionable than a wide confidence interval. In this error assessment task, JAWS methods achieved higher AUROC scores than split-conformal baselines, proving superior at identifying specific predictions likely to be erroneous.

Illustration of the Error Assessment task, which estimates the probability that a prediction stays within a user-defined safety tolerance.

Limitations and Future Directions

While JAWS offers robust theoretical guarantees, it has practical limitations. The current theory for JAW assumes access to "oracle weights"—the exact likelihood ratio defining the shift. In practice, this shift must be estimated from data, introducing potential estimation error. Furthermore, under extreme covariate shift, the normalized weights can approach the significance level $\alpha$, leading to uninformative, wide intervals. Finally, importance weighting naturally increases the variance of the estimates; future work will likely focus on variance reduction techniques to produce even sharper predictive intervals.

Summary updated

Abstract

We propose \textbf{JAWS}, a series of wrapper methods for distribution-free uncertainty quantification tasks under covariate shift, centered on the core method \textbf{JAW}, the \textbf{JA}ckknife+ \textbf{W}eighted with data-dependent likelihood-ratio weights. JAWS also includes computationally efficient \textbf{A}pproximations of JAW using higher-order influence functions: \textbf{JAWA}. Theoretically, we show that JAW relaxes the jackknife+'s assumption of data exchangeability to achieve the same finite-sample coverage guarantee even under covariate shift. JAWA further approaches the JAW guarantee in the limit of the sample size or the influence function order under common regularity assumptions. Moreover, we propose a general approach to repurposing predictive interval-generating methods and their guarantees to the reverse task: estimating the probability that a prediction is erroneous, based on user-specified error criteria such as a safe or acceptable tolerance threshold around the true label. We then propose \textbf{JAW-E} and \textbf{JAWA-E} as the repurposed proposed methods for this \textbf{E}rror assessment task. Practically, JAWS outperform state-of-the-art predictive inference baselines in a variety of biased real world data sets for interval-generation and error-assessment predictive uncertainty auditing tasks.
BibTeX
@inproceedings{prinster2022jaws,
  author = {Andrew Prinster and Anqi Liu and Suchi Saria},
  title = {{JAWS: Auditing Predictive Uncertainty Under Covariate Shift}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2022},
  url = {https://openreview.net/forum?id=TYMGhqlSFkC},
}
Figure