---

name: pgmpy description: "Route pgmpy causal and probabilistic graphical-model tasks to focused modeling, learning, inference, causal, data I/O, and extension workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

pgmpy

Use this repo skill when the task names pgmpy or asks for Python workflows around causal graphs, Bayesian networks, Markov networks, conditional probability distributions, structure learning, probabilistic inference, causal identification/effect estimation, example graphical models, model file formats, or pgmpy extension work.

pgmpy is a Python package for causal and probabilistic reasoning with graphical models. It provides graph/model classes, CPDs and factors, causal discovery, parameter estimation, inference, sampling/simulation, causal identification and prediction, datasets/example models, metrics, and read/write utilities.

First checks

• Install/use pgmpy with Python 3.10-3.14. For package use, start with pip install pgmpy or conda install conda-forge::pgmpy.
• For local repository maintenance, prefer pip install -e .[tests] and focused pytest targets. Do not commit or push from an agent session.
• Optional surfaces require explicit extras: pgmpy[torch] for FunctionalBayesianNetwork/FunctionalCPD; pgmpy[optional] for plotting, xgboost, and LLM-assisted discovery dependencies. Provider credentials/network are still separate requirements.
• To verify an installed package without source checkout access, run python scripts/check_pgmpy_environment.py --json from the skill root. Add --check-optional only when optional imports matter.
• Read references/repo-provenance.md before refreshing this skill for a changed checkout.
• Read references/package-map.md when choosing between canonical modules, legacy compatibility imports, optional extras, and public entry points.
• Read references/troubleshooting.md for cross-cutting install/import, optional dependency, state-name, network/cache, and package-version issues before diving into a sub-skill.

Route by task

User task	Read
Build a DAG, PDAG, ADMG, MAG, BN, DBN, Markov network, factor graph, CPD, or factor; debug check_model() or CPD shapes.	modeling-and-factors
Learn graph structure, choose PC/GES/HillClimb/Chow-Liu/TAN, select CI tests or scores, use expert knowledge, or fit CPDs from data.	learning-structure-and-parameters
Query posterior probabilities, run MAP, use Variable Elimination or Belief Propagation, sample/simulate from a fitted model, or work with DBN inference.	inference-sampling-and-simulation
Identify adjustment/frontdoor sets, use CausalInference.query(..., do=...), estimate ATE, or use causal prediction regressors.	causal-identification-and-effects
Load/list built-in datasets or example models, save/load BIF/XMLBIF/NET/UAI/XDSL files, or compute graph/data evaluation metrics.	data-io-and-evaluation
Add a new causal discovery algorithm, CI test, score, metric, dataset, or example model to a pgmpy checkout.	extending-pgmpy

Routing rules and boundaries

1. If the user only needs a graph object with roles and no CPDs, start in modeling-and-factors; do not force a Bayesian-network model.
2. If the user has data and wants a learned graph or learned CPDs, route to learning-structure-and-parameters, then back to modeling-and-factors only for model-family construction details.
3. If the user asks for P(Y | X=x), posterior marginals, MAP, or sampling under evidence, route to inference-sampling-and-simulation. If they ask for do(X=x), ATE, adjustment, frontdoor, instruments, or treatment effects, route to causal-identification-and-effects.
4. If a workflow begins by loading bnlearn/alarm, a dataset name, or a .bif/.uai/.xmlbif file, route first to data-io-and-evaluation, then to the workflow that consumes the loaded object.
5. Prefer canonical packages for new code: pgmpy.causal_discovery, pgmpy.parameter_estimator, pgmpy.structure_score, pgmpy.ci_tests, pgmpy.metrics, pgmpy.datasets, and pgmpy.example_models. Treat pgmpy.estimators as a legacy compatibility surface unless the task is explicitly maintaining backwards compatibility.
6. Do not treat optional torch/Pyro functional models, LLM-assisted discovery, plotting backends, or remote HuggingFace-backed assets as verified unless the user environment explicitly has the needed extras, credentials, network/cache, and hardware.

Common package facts

• Core CPU workflows use NetworkX, NumPy, SciPy, pandas, scikit-learn, statsmodels, tqdm, pyparsing, joblib, opt_einsum, scikit-base, and HuggingFace Hub helpers.
• BayesianNetwork and MarkovNetwork are deprecated aliases for DiscreteBayesianNetwork and DiscreteMarkovNetwork.
• Most user-facing algorithms expose sklearn-like constructors and fit(...); fitted discovery results are typically available as causal_graph_ and adjacency_matrix_.
• Model validation is not optional. Run check_model() before inference, simulation, serialization, or causal queries on parameterized models.
• Use show_progress=False in automated examples/tests unless progress bars are explicitly useful.

Safe validation

For a quick installed-package check:

python scripts/check_pgmpy_environment.py --json

For workflow-specific checks, use the nearest bundled script:

• modeling-and-factors/scripts/check_modeling_smoke.py
• learning-structure-and-parameters/scripts/learn_structure_smoke.py
• inference-sampling-and-simulation/scripts/inference_smoke.py
• causal-identification-and-effects/scripts/causal_effect_smoke.py
• data-io-and-evaluation/scripts/data_io_smoke.py
• extending-pgmpy/scripts/extension_template_check.py

These scripts are deterministic, no-network by default, and do not require the original repository checkout except extension_template_check.py, which is intentionally a checkout-maintenance helper and requires --repo when run away from the checkout.