Skip to content

README.md

Latest commit

 

History

History
144 lines (106 loc) · 4.25 KB

README.md

File metadata and controls

144 lines (106 loc) · 4.25 KB

SP-CCI: Synthetic-Powered Conformal Counterfactual Inference

This repository contains the code to run experiments for the paper:

Synthetic Counterfactual Labels for Efficient Conformal Counterfactual Inference

SP-CCI augments Conformal Counterfactual Inference (CCI) with synthetic counterfactual labels while preserving finite-sample coverage via a debiased, risk-controlled calibration procedure inspired by Prediction-Powered Inference (PPI) and Risk-Controlling Prediction Sets (RCPS).

Key features

  • Baseline CCI implementation.
  • SP-CCI with synthetic label augmentation and debiased miscoverage estimator.
  • Experiments on Synthetic data (Lei–Candès setup) and IHDP; optional wireless handover (Sionna RT) scaffold.
  • Figure scripts to reproduce plots like Coverage and Prediction Width.

Repository Structure

sp-cci/
├── pyproject.toml
├── requirements.txt
├── requirements-rt.txt
├── pre-commit-config.yaml
├── README.md
├── LICENSE
├── .gitignore
├── data/
│   └── README.md
├── results/
│   └── .gitkeep
├── outputs/
│   └── .gitkeep
├── Figures/
│   └── .gitkeep
├── scripts/
│   ├── run_synthetic.py
│   ├── run_ihdp.py
│   ├── make_figures.py
│   └── utils.py
└── sp_cci/
    ├── __init__.py
    ├── cci.py
    ├── spcci.py
    ├── metrics.py
    ├── data_synthetic.py
    ├── data_ihdp.py
    └── plotting.py
  • sp_cci/: library code (CCI, SP-CCI, data generators, metrics)
  • scripts/: CLIs to run experiments & render figures
  • data/: place external datasets here (IHDP, etc.) — see instructions below
  • results/, outputs/, Figures/: experiment logs, tables, and plots

Installation

1) Create environment

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -U pip
pip install -e .

Optional (wireless ray-tracing experiment scaffold):

pip install -r requirements-rt.txt

2) Pre-commit hooks (recommended)

pip install pre-commit
pre-commit install

Datasets

Synthetic (no download needed)

Generated on-the-fly

IHDP (semi-synthetic)

IHDP Dataset Source

This repo uses the widely adopted semi-synthetic IHDP dataset. Please cite the original study and the commonly used semi-synthetic construction:

  • Hill, J. (2011). Bayesian nonparametric modeling for causal inference. Journal of Computational and Graphical Statistics, 20(1), 217–240.
  • Shalit, U., Johansson, F. D., & Sontag, D. (2017). Estimating individual treatment effect: Generalization bounds and algorithms. In ICML 2017. (NPCI/CFR version of IHDP used broadly in CATE literature.) Place the processed IHDP table(s) under:
data/ihdp/ihdp.csv

You can adapt sp_cci/data_ihdp.py to your specific file name/format.

The code does not auto-download IHDP. Ensure you have rights to use the dataset.

Wireless (optional; Sionna RT)

The wireless demo is scaffolded only. It requires TensorFlow + Sionna and a compatible CUDA setup.

Usage

1) Synthetic experiment

python scripts/run_synthetic.py   --n 5000 --rho 0.0 --alpha 0.15 --delta 0.1   --runs 50 --quality HQ   --out results/synth_hq.json

2) IHDP experiment

python scripts/run_ihdp.py   --alpha 0.15 --delta 0.1 --runs 50   --quality MQ   --ihdp-path data/ihdp/ihdp.csv   --out results/ihdp_mq.json

3) Make figures

python scripts/make_figures.py   --inputs results/synth_hq.json results/ihdp_mq.json   --outdir Figures

This will produce PDFs like Coverage.pdf and prediction_width.pdf.

Methods (short)

  • CCI (weighted conformal prediction) constructs intervals for (Y(1)) using treated calibration data and propensity weighting.
  • SP-CCI augments treated calibration with synthetic counterfactual labels (\hat{Y}(1)) from a learned generator; a PPI-style debiasing step combined with RCPS selects the minimal widening (\eta) s.t. an upper confidence bound on miscoverage is ≤ (�lpha).

Citation

If this repository helps your research, please cite the paper.

License

MIT. See LICENSE.