False Coverage Proportion Control for Conformal Prediction
Authors: Alexandre Blain, Bertrand Thirion, Pierre Neuvial
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experimental validation on Open ML datasets, we demonstrate that our proposed methods effectively control the FCP and produce sharp prediction intervals. We use 17 Open ML (Vanschoren et al., 2014) datasets from (Grinsztajn et al., 2022). Each dataset is randomly split (nsplit = 30 times) into a train, calibration and test set. |
| Researcher Affiliation | Academia | 1INRIA 2Université Paris-Saclay 3Institut de Mathématiques de Toulouse, Université de Toulouse; CNRS; UPS, F-31062 Toulouse Cedex 9, France 4CEA. Correspondence to: Alexandre Blain <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Sampling order statistics of conformal pvalues using Proposition 1. Algorithm 2 Computing the Empirical JER. Algorithm 3 Performing calibration on conformal pvalues. |
| Open Source Code | Yes | An implementation of Co JER is available at https://github.com/sanssouci-org/Co JER-paper, together with the code to reproduce the numerical results of this paper. |
| Open Datasets | Yes | We use 17 Open ML (Vanschoren et al., 2014) datasets from (Grinsztajn et al., 2022). |
| Dataset Splits | No | Each dataset is randomly split (nsplit = 30 times) into a train, calibration and test set. The paper specifies the creation of train, calibration, and test sets and the number of splits (30 times) but does not provide explicit proportions (e.g., 70/15/15) for these splits. |
| Hardware Specification | Yes | All experiments were performed using 40 CPUs, Intel(R) Xeon(R) CPU E5-2660 v2 @ 2.20GHz |
| Software Dependencies | No | The paper mentions models like Random Forest, Multi-Layer Perceptron, Support Vector Regression, K-Nearest Neighbors, and Lasso, but does not provide specific version numbers for any software libraries or frameworks used (e.g., Python, scikit-learn, PyTorch versions). |
| Experiment Setup | Yes | We use α = 0.1 for all methods. For FCP controlling methods, we set δ = 0.1 and use SCP with the largest level α such that FCPα ,δ α. |