windisch/bn_testing

A test bench for graphical models

BN testing

A test framework to evaluate methods that learn Bayesian Networks from high-dimensional observational data.

Sampling

Set up the graphical model and sample data

from bn_testing.models import BayesianNetwork
from bn_testing.dags import ErdosReny
from bn_testing.conditionals import PolynomialConditional


model = BayesianNetwork(
   dag=ErdosReny(p=0.01, n_nodes=100),
   conditionals=PolynomialConditional(max_terms=5)
)

df = model.sample(10000, normalize=True)

The observations are stored in a pandas.DataFrame where the columns are the nodes of the DAG and each row is an observation. The underlying DAG of the graphical model can be accessed with model.dag

bn_testing's Issues

Allow variable transforms in terms

Add parameter that allows to transforms a (selection?) of the variables before applying a term.

For instance, it should be possible in the term Polynomial(["X", "Y"], [[1, -1]], [1]) that the the variable Y is replaced with exp(Y) before the term is applied to avoid divisions by zero errors.

How about specifying nodes_per_group and avoid that size of the last group is uncontrolled ?

How about specifying nodes_per_group and avoid that size of the last group is uncontrolled ?

Originally posted by @azeller14 in #1 (comment)

Allow to save and load models

Allow discrete variables

Allow to add hidden nodes

Allow modification of source nodes