erdogant/bnlearn

bnlearn – Python library for Bayesian‑network learning & causal analysis

What it does – bnlearn lets you build, train, and query probabilistic graphical models (Bayesian networks). It covers the whole pipeline:

• Structure learning – discover the directed‑acyclic graph (DAG) that best explains your data, using classic scores (BIC, K2, BDeu) and search strategies (exhaustive, hill‑climbing, Chow‑Liu, TAN, etc.).
• Parameter learning – estimate conditional probability tables (CPTs) for a given DAG, even with incomplete data.
• Inference & prediction – compute posterior, interventional (do‑calculus) and counter‑factual queries, and make causal predictions.
• Synthetic data & utilities – sample new data from a learned model, discretise continuous variables, impute missing values, and convert between graph representations (adjacency matrix, edge list, Graphviz).
• Visualization – plot the network structure with Matplotlib or Graphviz, compare two networks, and export models.

All of this is wrapped in a tidy, high‑level API (bn.structure_learning.fit(), bn.parameter_learning.fit(), bn.inference.fit(), bn.predict(), …) so you can go from a raw pandas DataFrame to a fully‑fledged causal model in a few lines of code.

---

Quick start

# Install (Python 3.10+ required)
pip install bnlearn          # or via conda, uv, or directly from GitHub

import bnlearn as bn

# Load an example dataset (the classic sprinkler network)
df = bn.import_example()

# 1️⃣ Learn the network structure from data
model = bn.structure_learning.fit(df)

# 2️⃣ Optionally compute edge‑strength statistics
model = bn.independence_test(model, df)

# 3️⃣ Visualise the learned DAG
bn.plot(model)

# 4️⃣ Learn CPTs (parameter learning)
model = bn.parameter_learning.fit(model, df)

# 5️⃣ Perform causal inference
query = bn.inference.fit(model, variables=['Rain'],
                         evidence={'Cloudy': 1, 'Sprinkler': 0, 'Wet_Grass': 1})
print(query.df)

The library also provides ready‑made pipelines for specific tasks (e.g., bn.sampling() to generate synthetic data, bn.knn_imputer() for missing‑value imputation, bn.discretize() for binning continuous features).

---

Where it fits

• Causal discovery – ideal for researchers and data scientists who need to uncover directed relationships from observational data.
• Probabilistic modelling – useful for any domain where uncertainty must be quantified (medicine, finance, reliability engineering, etc.).
• Education & prototyping – the extensive documentation, notebooks, and blog‑post tutorials make it a good teaching tool for Bayesian‑network concepts.

---

Documentation & community

• Full API docs – https://erdogant.github.io/bnlearn/
• Example notebooks – Google‑Colab links from the docs.
• Blog & podcasts – medium articles and short audio explainers for each major feature.
• Open source – MIT‑licensed, actively maintained (≈ 2 k stars, regular releases), and welcomes contributions.

---

TL;DR

bnlearn is a well‑documented, pip‑installable Python package that streamlines the end‑to‑end workflow of Bayesian‑network based causal modelling: discover structure, learn parameters, run inference, generate data, and visualise results—all with a few high‑level function calls.