py-pcalg

Lightweight PC algorithm for causal skeleton discovery — pip + CLI + graph viz from a CSV in ~30 seconds.

纯 Python PC 算法（Peter–Clark）：从观测数据画因果骨架。比 R pcalg 更轻，比 causal-learn 更易读。

pip install -e .
pcalg demo
# or: pcalg run your_data.csv --alpha 0.05 -o skeleton.png

Useful? Star ⭐ so others can find a small causal-discovery CLI.

---

Why this tool

	py-pcalg	causal-learn	R pcalg
Install	pip + CLI	heavier deps	R runtime
Code size	~1k LOC, readable	full toolkit	reference impl
Use case	quick skeleton + viz	research pipeline	production stats

---

Install

pip install -e .          # from source
# after PyPI publish:
# pip install py-pcalg

Dependencies: numpy, pandas, matplotlib, networkx.

---

CLI

# bundled demo (test_data.csv)
pcalg demo

# your CSV (numeric columns)
pcalg run your_data.csv --alpha 0.05 -o skeleton.png

# headless CI
pcalg run your_data.csv --no-plot

---

Python API

import pandas as pd
from pcalg import skeleton, independence_test, visualize_graph

data = pd.read_csv("your_data.csv")
corr = data.corr()
labels = list(range(data.shape[1]))
label_dict = {i: str(c) for i, c in enumerate(data.columns)}

graph = skeleton(
    suff_stat=[corr.to_numpy(), len(data)],
    indep_test=independence_test,
    alpha=0.05,
    labels=labels,
)

visualize_graph(graph, label_dict, save_path="skeleton.png")
print(graph.M)

Legacy imports still work: from pc import skeleton, from utils import Matrix.

---

Algorithm

1. Start from a complete undirected graph.
2. Skeleton discovery: for each edge (X, Y), search conditioning sets S; remove X–Y if X ⊥ Y | S (Fisher-Z test, α = 0.05 by default).
3. Orientation (partial): V-structures and Meek rules — full CPDAG orientation is a roadmap item.

Independence test: Fisher's Z on partial correlations (see pcalg.utils.independence_test).

---

Project layout

py_pcalg/
├── pcalg/           # installable package
│   ├── pc.py        # skeleton(), pc()
│   ├── utils.py     # Matrix, Fisher-Z tests
│   ├── graph.py     # NetworkX + matplotlib viz
│   └── cli.py       # pcalg command
├── tests/           # pytest
├── test_data.csv    # demo dataset
└── pyproject.toml

---

Tests

pip install -e ".[dev]"
pytest

---

Related work

• causal-learn — CMU toolkit (PC, FCI, GES, …)
• gCastle — Huawei causal structure learning
• pcalg (R) — original reference implementation

References: Spirtes, Glymour & Scheines (2000); Kalisch & Bühlmann (2007).

---

Assumptions & limits

• Causal Markov + faithfulness; no latent confounders (use FCI for that).
• Needs sufficient sample size for stable partial-correlation tests.
• Dense graphs can be slow — m_max caps conditioning-set size.

---

License

MIT — see repository root.