This package provides some statistical wrappers for machine learning classification tools in order to construct prediction sets for the label of a new test point with provably valid marginal coverage and approximate conditional coverage.

Accompanying paper (https://papers.nips.cc/paper/2020/hash/244edd7e85dc81602b7615cd705545f5-Abstract.html):

"Classification with Valid and Adaptive Coverage"
Y. Romano, M. Sesia, E. Candès
NeurIPS 2020 (spotlight).

• arc/ Python package implementing our methods and some alternative benchmarks.
• third_party/ Third-party Python packages imported by our package.
• examples/ Jupyter notebooks with introductory usage examples.
• experiments_sim_data Code for the experiments with simulated data discussed in the accompanying paper.
• experiments_real_data Code for the experiments with real data discussed in the accompanying paper.

This package builds upon the following non-standard Python packages provided in the "third-party" directory:

• nonconformist https://github.com/donlnz/nonconformist
• cqr https://github.com/yromano/cqr
• cqr-comparison https://github.com/msesia/cqr-comparison

Prerequisites for the arc package:

• numpy
• scipy
• sklearn
• skgarden
• torch
• tqdm

Additional prerequisites for example notebooks:

• pandas
• matplotlib
• seaborn

The development version is available from GitHub:

git clone https://github.com/msesia/arc.git

This project is licensed under the MIT License - see the LICENSE file for details.