Conformal Prediction with Model-Aware Debiasing
Keywords: conformal prediction, model-aware debiasing, statistical inference, prediction interval
Abstract: Bias in model estimation can lead to wider prediction intervals, diminishing the utility of predictive inference. Existing methods have attempted to address this issue, but they often rely on nontrivial assumptions such as specific error distributions or model sparsity, and fail to guarantee coverage in finite samples, which makes their predictions unreliable in practice. To overcome these limitations, we propose a model-aware conformal prediction method that utilizes known model information to achieve debiasing while leaving the unknown aspects, such as data distribution, to the conformal prediction framework. This approach requires only the assumption of exchangeability, making it broadly applicable across various models. Importantly, it retains the finite-sample coverage property and produces shorter prediction intervals compared to existing methods. When applied to threshold ridge regression, we theoretically demonstrate that the model-aware conformal prediction maintains finite-sample marginal coverage and, under certain assumptions, converges to the oracle prediction band, achieving asymptotic conditional validity. Numerical experiments further show that our method outperforms existing methods, providing more efficient prediction intervals across diverse regression datasets.
Supplementary Material: zip
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8376
Loading