Reproduce “Training-Conditional Coverage Bounds under Covariate Shift” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mehrdad Pournaderi, Yu Xiang
Researchers at the University of Utah established finite-sample Probably Approximately Correct (PAC) training-conditional coverage bounds for weighted conformal prediction methods, including split conformal, full conformal, and jackknife+, when operating under covariate shift. These bounds quantify how the severity of the distribution shift and the algorithmic stability impact the reliability of prediction sets.
Key insights
- The concentration and quality of training-conditional coverage are directly influenced by the severity of covariate shift, as characterized by the statistical properties of the likelihood ratio dQ/dP.
- Algorithmic stability, particularly uniform stability, is crucial for achieving strong training-conditional coverage guarantees for methods like jackknife+ and full conformal under covariate shift.
- For models satisfying uniform stability, the derived bounds demonstrate significantly better sample complexity rates for training-conditional coverage (e.g., O(n^-1/2)) compared to bounds based on weaker stability notions like (m,n)-stability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.