Multimodal Bayesian-network example

This example uses the package's deterministic teaching network with Condition, Gaze, Pupil, EDA, Trust, and Choice. The variables are discrete only so the full probability logic is easy to inspect; this is not a recommendation to dichotomize continuous physiology.

Generate reproducible demonstration data

from eyeprocesspy.bayesian_networks import (
    simulate_multimodal_bayesian_network_example,
    prepare_bayesian_network_data,
    define_bn_constraints,
    learn_bayesian_network,
    fit_bayesian_network,
    query_bayesian_network,
)

data = simulate_multimodal_bayesian_network_example(
    n_participants=1000,
    random_state=2026,
)

bn_data = prepare_bayesian_network_data(
    data,
    nodes={
        "condition": {"type": "categorical", "modality": "experimental"},
        "gaze": {"type": "categorical", "modality": "gaze"},
        "pupil": {"type": "categorical", "modality": "pupil"},
        "eda": {"type": "categorical", "modality": "physiology"},
        "trust": {"type": "categorical", "modality": "questionnaire"},
        "choice": {"type": "categorical", "modality": "behavior"},
    },
    participant_id="participant_id",
    trial_id="trial_id",
)

constraints = define_bn_constraints(
    temporal_tiers=[
        ["condition"],
        ["gaze"],
        ["pupil", "eda"],
        ["trust"],
        ["choice"],
    ],
    nodes=bn_data.structure_nodes,
    max_parents=3,
)

Fit a fixed theoretical DAG

fitted = fit_bayesian_network(
    [
        ("condition", "gaze"),
        ("gaze", "pupil"),
        ("gaze", "eda"),
        ("condition", "trust"),
        ("gaze", "trust"),
        ("trust", "choice"),
    ],
    data=bn_data,
    estimator="bayesian",
    prior="BDeu",
    equivalent_sample_size=10,
)

Query evidence propagation

posterior = query_bayesian_network(
    fitted,
    target="trust",
    evidence={
        "condition": "Explanation",
        "pupil": "Dilated",
        "eda": "High",
    },
)

print(posterior.posterior)

The teaching probabilities were selected so that multimodal evidence raises the probability of high gaze inspection and subsequently high trust. The exact empirical posterior in the simulated dataset varies with sample size and seed because the table is sampled from the network.

Learn rather than fix the graph

learned = learn_bayesian_network(
    bn_data,
    algorithm="hill_climb",
    score="bic",
    constraints=constraints,
)

Do not read a recovered direction as a causal finding. Compare the learned graph with the known generating DAG using validate_bayesian_network_recovery() when doing simulation studies.