FAB-PPI: Frequentist, Assisted by Bayes, Prediction-Powered Inference

This repository contains code for FAB-PPI: Frequentist, Assisted by Bayes, Prediction-Powered Inference.

Requirements

The project dependencies are listed in the pyproject.toml file. We provide an additional environment.yaml file that can be used to create a Mamba (or Conda) environment with all the required dependencies to run the experiments:

    mamba env create -f environment.yaml && mamba activate fab_ppi-env

Experiments

The examples directory contains notebooks to replicate the experiments in the paper:

• simulation_biased.ipynb: Replicates the biased predictions study in Section 5.1
• simulation_noisy.ipynb: Replicates the noisy predictions study in Section 5.1
• alphafold.ipynb: Replicates the mean estimation experiment on the alphafold dataset mentioned in Section 5.2
• forest.ipynb: Replicates the mean estimation experiment on the forest dataset mentioned in Section 5.2
• galaxies.ipynb: Replicates the mean estimation experiment on the galaxies dataset mentioned in Section 5.2
• census_healtcare.ipynb: Replicates the logistic regression experiment on the healthcare dataset mentioned in Section 5.2
• gene_expression.ipynb: Replicates the quantile estimation experiment on the genes dataset mentioned in Section 5.2
• census_income.ipynb: Replicates the quantile estimation experiment on the census dataset mentioned in Section 5.2
• forest_horseshoe_scaled.ipynb: Replicates forest.ipynb with different horseshoe scales