PPCI: Prediction-Powered Conditional Inference

This repository contains the implementation of Prediction-Powered Conditional Inference (PPCI). PPCI enables valid uncertainty quantification for conditional functionals (e.g., conditional means) in settings with scarce labeled data, rich unlabeled data, and an available black-box predictor.

---

Requirements

• CUDA-enabled GPU environment
• Core computation is based on CuPy

---

Repository Structure

The project has been reorganized into two main independent directories based on the type of inference:

1. Conditional_mean_inference/
2. Conditional_quantile_inference/

Each directory contains its own core functions, simulation scripts, and real-data applications.

---

Part 1: Conditional Mean Inference

All files for mean inference are located in the Conditional_mean_inference/ directory.

Core File:

• conditional_mean_functions.py — base functions for conditional mean inference

1) Simulation Experiments

• Notebook: simu_mean.ipynb
• Outputs: CSV results are saved to ./results/ and PDFs are generated in the root of this module.

2) Census Income Data Analysis

• Location: ./census_income_data/
• Notebooks to run: - Income_data_mean.ipynb
  • Income_data_different_unlabeled_sample_size.ipynb
• Outputs: CSV results and PDFs are saved to ./census_income_data/results/

3) BlogFeedback Data Analysis

• Location: ./blogfeedback_data/
• Important: run in the following order:
  1. Pre-processing: Run data_process.ipynb first. This step generates processed predictions, training data, and test data used by the main analysis.
  2. Main analysis: Run blogfeedback_data_mean.ipynb
• Outputs: CSV results and PDFs are saved to ./blogfeedback_data/results/

---

Part 2: Conditional Quantile Inference

All files for quantile inference are located in the Conditional_quantile_inference/ directory.

Core File:

• conditional_quantile_functions.py — base functions for conditional quantile inference

1) Simulation Experiments

• Notebook to run: Simu_conditional_quantile_new.ipynb
• Plotting (optional): plot_simu_qunatile.ipynb
• Outputs: CSV results are saved to ./results/ and PDFs (e.g., simu_quantile_rmse_coverage_width_3x3.pdf) are generated in the root of this module.

2) Census Income Data Analysis

• Location: ./census_income_data/
• Notebook to run: Income_data_conditional_quantile_new.ipynb
• Plotting (optional): income_data_plot.ipynb
• Outputs: CSV results and PDFs are saved to ./census_income_data/results/

---

Output Locations Summary

Outputs are handled independently within their respective inference folders:

• Simulation: [Inference_Type]/results/
• Census Income: [Inference_Type]/census_income_data/results/
• BlogFeedback (Mean only): Conditional_mean_inference/blogfeedback_data/results/