Installation

Installing bnlearn is straightforward. It is recommended to create a new environment for the installation.

conda create -n env_bnlearn python=3.8
conda activate env_bnlearn
pip install bnlearn

Quick Examples

Let’s start by importing some data. We need a DAG and CPD (Conditional Probability Distribution).

import bnlearn as bn

# Import example dataset
df = bn.import_example()

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

# Perform independence tests
model = bn.independence_test(model, df)

# Visualize the network
G = bn.plot(model)

Here’s another example demonstrating a complete workflow:

import bnlearn as bn

# Import a predefined DAG (Sprinkler network)
model = bn.import_DAG('sprinkler')

# Import example dataset
df = bn.import_example()

# Generate samples from the model
df = bn.sampling(model)

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

# Learn structure from data
model_sl = bn.structure_learning.fit(df)

# Learn parameters
model_pl = bn.parameter_learning.fit(model_sl, df)

# Compare networks
scores, adjmat = bn.compare_networks(model_sl, model)

Comparison between Bayesian Libraries

Feature	bnlearn	pgmpy	CausalNex	DoWhy	pyAgrum	CausalImpact
Primary Purpose	End-to-end Bayesian network modeling, structure learning, parameter learning, and inference	Low-level probabilistic graphical model toolkit	Causal discovery and reasoning (discrete data)	Causal inference with explicit treatment/outcome	General Bayesian/Markov model creation, inference, and visualization	Bayesian structural time-series for causal impact estimation
Structure Learning	[x] Multiple methods (Hillclimb, Constraint, Chow-Liu, Naive Bayes, TAN)	[x] Multiple estimators (HillClimb, PC, etc.	[x] NOTEARS algorithm (score-based)	[ ] Requires DAG as input	[x] Greedy or score-based approaches	[ ] Not applicable
Parameter Learning	[x] One-line CPD estimation	[x] Requires explicit setup	[x] Supported but requires preprocessing	[-] Limited; works with provided DAG	[x] Fully supported	[-] Limited; implicit (time-series reg.)
Inference / Do-calculus	[x] Built-in, intuitive syntax	[x] Variable elimination, sampling	[x] Supported (fitted CPDs)	[-] Treatment-outcome inference only	[x] Multiple algorithms	[x] Bayesian inference for intervention
Synthetic Data Generation	[x] Built-in (bn.sampling, bn.forward_sample)	[-] Possible manually	[ ] Not supported	[ ] Not supported	[x] Supported	[-] Possible via simulation only
Input Data Support	Discrete, continuous, or mixed (string labels allowed)	Discrete only	Discrete only	Encoded numeric (binary treatment)	Discrete or continuous	Continuous time series
Causal DAG Learning	[x] Fully unsupervised (independence tests, whitelist/blacklist edges)	[x] Requires explicit setup	[x] NOTEARS-based (numeric)	[ ] Requires predefined DAG	[x] Supported	[ ] Not applicable
Best For	[x] Practical causal modeling, exploratory discovery, education	[x] Developers needing flexibility	[x] Categorical causal discovery	[x] Econometric and policy impact estimation	[x] Education and simulation	[x] Intervention analysis and time-series
Ease of Use	Very high (auto-handles preprocessing, pipelines)	Moderate to low (manual setup required)	Moderate (needs encoding, tuning)	Low (requires causal knowledge)	Moderate	High (simple interface)
Learning Curve	Easy	Moderate	Moderate	Steep	Moderate	Easy
Visualization	[x] Static + interactive (Plotly/Graphviz)	[-] Basic plotting	[x] NetworkX-based graph	[-] Limited	[x] Interactive GUI	[-] Time-series plots only
Dependencies	Lightweight (bnlearn, pgmpy, pandas, numpy)	pgmpy, pandas, numpy	pgmpy, scikit-learn, networkx	dowhy, econml	aGrUM backend	CausalImpact, statsmodels, pandas
Causal Effect Estimation	[x] Via inference/intervention	[x] Via queries	[x] Via CPD inspection	[x] Core functionality (ATE, backdoor, IV)	[x] Via intervention modeling	[x] Core functionality (timeseries effects)
Preprocessing Effort	Low (automatic handling, independence testing)	Medium (manual setup)	High (encoding, state definition)	High (requires treatment/outcome encoding)	Medium	Low (requires clean time series)
Typical Use Cases	Exploratory causal discovery, inference, simulation, education	Custom Bayesian model design	Discrete causal discovery, NOTEARS DAGs	Estimating treatment effects (A/B tests)	Teaching, simulation	Program evaluation, A/B test
Advantages	End-to-end pipeline, supports string labels, whitelist/blacklist, independence tests, modern compatibility	Modular and flexible	NOTEARS algorithm, explicit DAG control	Rigorous causal framework, ATE estimation	Strong visualization tools	Simple time-series causality with Bayesian backbone
Disadvantages	Limited to moderate-scale DAGs	Verbose code, no auto-pipeline	No continuous data, high preprocessing	No structure learning, binary treatments	Verbose syntax	Only handles time-series, not generic DAGs

---

---