Defining a Custom Model#
When you already know the structure of a model, you can define the graph and its parameters directly in pgmpy. This workflow gives you full control over nodes, edges, and CPDs, and supports several model families beyond standard discrete Bayesian networks.
At a Glance#
Unified API: Create a model, add CPDs, and validate — the same pattern across all model families.
Multiple Model Families: Discrete, continuous, functional, dynamic, and undirected models.
Model Validation: Verify consistency of structure and parameters before downstream use.
API#
The common pattern is: choose a model class, define the structure, add CPDs, and validate:
from pgmpy.models import DiscreteBayesianNetwork
from pgmpy.factors.discrete import TabularCPD
model = DiscreteBayesianNetwork([("A", "B"), ("A", "C")])
cpd_a = TabularCPD("A", 2, [[0.6], [0.4]])
cpd_b = TabularCPD("B", 2, [[0.9, 0.2], [0.1, 0.8]], evidence=["A"], evidence_card=[2])
cpd_c = TabularCPD("C", 2, [[0.3, 0.7], [0.7, 0.3]], evidence=["A"], evidence_card=[2])
model.add_cpds(cpd_a, cpd_b, cpd_c)
print(model.check_model())
Once valid, the model works with inference, simulation, and the rest of the pgmpy stack.
Model Families#
pgmpy supports several model families, each paired with a matching CPD or factor type. These include discrete Bayesian networks with tabular CPDs, continuous networks with linear Gaussian CPDs, functional models with arbitrary CPDs, dynamic Bayesian networks for temporal processes, undirected Markov networks, and structural equation models.
Each family uses the same create → add CPDs → validate pattern shown above. For example, a continuous model:
from pgmpy.models import LinearGaussianBayesianNetwork
from pgmpy.factors.continuous import LinearGaussianCPD
model = LinearGaussianBayesianNetwork([("X", "Y"), ("Z", "Y")])
cpd_x = LinearGaussianCPD("X", beta=[0.0], std=1.0)
cpd_z = LinearGaussianCPD("Z", beta=[0.0], std=1.0)
cpd_y = LinearGaussianCPD("Y", beta=[0.2, 0.5, 0.3], std=1.0, evidence=["X", "Z"])
model.add_cpds(cpd_x, cpd_z, cpd_y)
print(model.check_model())
Model Validation#
model.check_model() verifies that the graph structure and attached CPDs are consistent
— correct parent sets, probability tables that sum to 1, matching cardinalities, etc.
Always validate before using a model downstream.
See Also#
See also
Parameter Estimation — Learn CPDs from data instead of defining them manually.
Probabilistic Inference — Query the model once it is built and validated.
API Reference#
For the full list of model classes and factor types: