RLCP

The project aims at developing a locally-weighted conformal prediction algorithm, which can attain meaningful local coverage guarantees in a model-free manner.

Overview

We provide the exact code for reproducing the numerical results in the paper "Conformal prediction with local weights: randomization enables robust guarantees".

Folders

• utils/: contains the helper functions for all the experiments.
• simulations/: contains the codes for simulation experiments in the paper.
• real_data/: contains the codes for real data experiments in the paper.
• results/: contains the results from the real data and simulation experiments.

Guide for the codes in simulations/ folder

Codes for reproducing results in Section 5.1

• marginal coverage (Figure 1): mc_cov.R/
• average prediction interval and local coverage (Figure 2,10 & 11): 1d_local_coverage.R/
• marginal coverage on subsets with constant bandwidth (Figure 3): mc_onA_const_h.R/
• marginal coverage on subsets with dimension-adaptive bandwidth (Figure 4): mc_onA_opt_h.R/

Other simulation results from Section 6 and appendix

• effect of randomization on RLCP prediction interval in simulation settings (Figure 7 & 8 left): simul_rand_effect.R/
• coverage of mRLCP prediction interval (Figure 9): mRLCP.R/
• comparing RLCP with Gibbs et al. (Figure 12): Conformal_comparison.ipynb and RLCP_Gibbs_comparison.R
• local coverage on grids (Figure 13): grid_local_coverage.R

Guide for the codes in real_data/ folder

Codes for reproducing results in Section 5.2

• marginal coverage and conditional coverage (Figure 5 & 6): real_exp.R/
• quantitative analysis on effect of randomization on RLCP prediction interval in real data experiments (Figure 8 right): real_dev.R/

Real Data

For our real data experiments, we have used the Train_Data.csv of medical insurance dataset from kaggle datasets. For reproducing the real data experiments, please download the data from this above link and load it into your R before running the codes from real_data/ folder.