Conformalized Unconditional Quantile Regression

Ahmed M. Alaa,Zeshan Hussain,David Sontag

2023-04-04

Abstract:We develop a predictive inference procedure that combines conformal prediction (CP) with unconditional quantile regression (QR) -- a commonly used tool in econometrics that involves regressing the recentered influence function (RIF) of the quantile functional over input covariates. Unlike the more widely-known conditional QR, unconditional QR explicitly captures the impact of changes in covariate distribution on the quantiles of the marginal distribution of outcomes. Leveraging this property, our procedure issues adaptive predictive intervals with localized frequentist coverage guarantees. It operates by fitting a machine learning model for the RIFs using training data, and then applying the CP procedure for any test covariate with respect to a ``hypothetical'' covariate distribution localized around the new instance. Experiments show that our procedure is adaptive to heteroscedasticity, provides transparent coverage guarantees that are relevant to the test instance at hand, and performs competitively with existing methods in terms of efficiency.

Machine Learning,Methodology

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to provide highly adaptable and transparent prediction intervals in predictive inference. Specifically, the authors propose a method that combines Conformal Prediction (CP) with Unconditional Quantile Regression (UQR) to construct prediction intervals that can adapt to the uncertainty of different prediction instances and provide coverage guarantees relevant to each specific instance. Traditional methods are often only able to provide marginally effective fixed - length prediction intervals, which may not be informative or accurate enough for specific prediction instances. Therefore, the main objective of this paper is to develop a predictive inference method that can provide conditional coverage effectiveness while maintaining marginal coverage effectiveness, that is, it can adjust the length of the prediction interval according to the specific uncertainty of the prediction instance and report the coverage guarantee related to each prediction instance.

The main contributions in the paper can be summarized as follows:

1. **Highly adaptable prediction intervals**: By combining CP and UQR, a method is proposed that can adaptively adjust the length of the prediction interval among different prediction instances, thereby better reflecting the uncertainty of each instance.

2. **Transparent coverage guarantees**: Compared with traditional marginal coverage guarantees, this method can provide conditional coverage guarantees relevant to each prediction instance, improving the transparency and credibility of the prediction results.

3. **Effectiveness in local regions**: This method not only provides effective coverage guarantees globally, but also can provide effective coverage guarantees within local regions (that is, within the set of instances similar to the new test point), further enhancing the adaptability and accuracy of the prediction.

4. **Theoretical guarantees**: The paper provides a theoretical analysis and proves that the proposed CUQR method can achieve a high - probability conditional coverage guarantee under certain conditions, that is, within each relevant subgroup, the coverage rate of the prediction interval is close to the preset target value.

Through the above contributions, this paper aims to improve the existing predictive inference methods, making them more suitable for application scenarios that require precise quantification of uncertainty and the provision of transparent coverage guarantees, such as clinical decision - support systems.