Multi-Source Conformal Inference Under Distribution Shift

Authors: Yi Liu, Alexander Levis, Sharon-Lise Normand, Larry Han

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate our proposed method by conducting extensive Monte Carlo simulations, examining aspects such as marginal coverage, conditional coverage, and the width of the prediction interval. [...] Hospital length of stay prediction intervals for pediatric patients undergoing a high-risk cardiac surgical procedure between 2016-2022 in the U.S. illustrate the utility of our methodology.
Researcher Affiliation Academia 1North Carolina State University, Department of Statistics, Raleigh, NC, USA 2Carnegie Mellon University, Department of Statistics, Pittsburgh, PA, USA 3Harvard Medical School, Department of Health Care Policy, Boston, MA, USA 4Northeastern University, Department of Health Sciences, Boston, MA, USA.
Pseudocode Yes Algorithm 1 Robust multi-source conformal prediction
Open Source Code Yes We provide a user-friendly R function Mu SCI() implementing the proposed method with an illustrative example, available at: https://github.com/yiliu1998/Multi-Source-Conformal.
Open Datasets Yes We utilize data from the Society of Thoracic Surgeons Congenital Heart Surgery Database (STS-CHSD) which includes audited preoperative, intraoperative, and early postoperative information (Overman et al., 2019) from U.S. congenital heart surgery centers.
Dataset Splits Yes Split the training data D randomly into D1 and D2, where Dj = {Oi D, i Ij} for j = 1, 2 and I1 I2 = {1, 2, . . . , n}. [...] We perform cross-fitting such that the nuisance estimators ( bm, bη, bq0) are estimated on an independent data split from the given estimating equation. [...] λ is a tuning parameter chosen by cross-validation.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions software like Super Learner, random forest, elastic net, GLM, and R, but does not provide specific version numbers for these software components.
Experiment Setup Yes In total, we consider 3 sample sizes (300, 1000, 3000) 3 levels of covariate shift (homogeneous, weakly heterogeneous, strongly heterogeneous) 2 types of outcome errors (homoskedastic, heteroskedastic) 3 levels of concept shift (CCOD holds, weak violation, strong violation) 3 different conformal scores (ASR, locally weighted ASR, CQR) = 162 scenarios for our proposed method and the five competitor methods. [...] We generate data from K = 5 sites, where site 0 is the target site and sites 1 through 4 are source sites, and Ti {0, , 4} denotes the site of subject i. Our goal is to construct valid prediction intervals for a testing point from the target site. We consider the sample size in each site to be nk {300, 1000, 3000}, k = 0, ..., 4 and generate data over M = 500 independent Monte Carlo replications. [...] λ is a tuning parameter chosen by cross-validation.