Probabilistic Inference

Use variables, CPDs, factor tables, and evidence when graph records contain uncertainty. TongGraph compiles an active subgraph and runs finite discrete belief propagation locally.

from tonggraph import Graph

graph = Graph()

sensor = graph.add_node("sensor:1", labels=["Sensor"])
alert = graph.add_node("alert:1", labels=["Alert"])
graph.add_edge(sensor, alert, "INFLUENCES")

sensor_ok = graph.add_variable(
    "binary",
    owner_id=sensor,
    prior={"p": 0.95},
)
alert_active = graph.add_variable(
    "binary",
    owner_id=alert,
    prior={"p": 0.05},
)

graph.add_cpd(
    alert_active,
    [sensor_ok],
    [
        0.8,
        0.2,
        0.1,
        0.9,
    ],
)

active = graph.compile_active_subgraph(
    [alert_active],
    evidence={sensor_ok: "false"},
    radius=1,
)
result = graph.belief_propagation(
    [alert_active],
    evidence={sensor_ok: "false"},
    radius=1,
    persist=True,
)

print(active)
print(result["converged"], result["max_residual"])
print(graph.posterior(alert_active))

Belief propagation is approximate on loopy graphs. Use the returned convergence metadata and traces when interpreting results.