Add causal inference: bn.inference.fit(..., do={...}) (closes #79)

[dbolser]

Adds Pearl's do-operator to bn.inference.fit, so interventions are available from Python the way mutilated() is in the R version of bnlearn. Closes #79.

model = bn.import_DAG('sprinkler')
bn.inference.fit(model, variables=['Wet_Grass'], evidence={'Sprinkler': 1})  # P(W=1|S=1)    = 0.927
bn.inference.fit(model, variables=['Wet_Grass'], do={'Sprinkler': 1})        # P(W=1|do(S=1))= 0.945

Observing Sprinkler=1 is evidence that it was probably not cloudy; setting it by intervention says nothing about the weather. do combines freely with evidence, and interventions are labeled do(Sprinkler)=1 in the readable summary.

How it works, and why not CausalInference

The obvious implementation is pgmpy's CausalInference.query, but its adjustment-based estimator gives wrong answers on two cases this feature needs, both verified against hand-derived CPT values on pgmpy 1.1.2 and on current pgmpy dev (reported as pgmpy/pgmpy#3543):

query	exact	CausalInference.query
P(W=1 | do(S=1), R=1) — evidence outside the adjustment set	0.99	0.9612
P(C=1 | do(A=1), do(B=1)) — interventions on causally related nodes	1.0	0.7

So do instead runs VariableElimination over model.do(...)'s mutilated network with the intervened variables fixed as evidence — in a network whose incoming edges are cut, conditioning is intervening. That is exact in both cases above, and it keeps elimination_order and joint working with do. If pgmpy fixes #3543, nothing here needs to change.

Notes

• do is the last parameter of fit, so existing positional calls are unaffected.
• Validation: unknown node names in do, and a variable given in both do and evidence, raise instead of silently misbehaving.
• Also fixes a pre-existing crash: evidence=None raised AttributeError on the evidence.keys() validation line.
• 4 new tests (locked values hand-derived from the CPDs, incl. both cases in the table); full suite 94/94 green locally under pgmpy 1.1.2, and green on all 9 CI jobs (Python 3.10/3.12/3.13 × ubuntu/macos/windows) on the fork.

🤖 Generated with Claude Code