[Repo Assist] perf(falsify): replace N×nx.ancestors calls with transitive_closure_dag in validate_pd

[github-actions]

🤖 This PR was created by Repo Assist, an automated AI assistant.

Summary

validate_pd in dowhy/gcm/falsify.py was building the list of (ancestor, node) pairs by calling nx.ancestors() once per graph node:

# Before — O(N·(V+E)): N separate BFS/DFS traversals
pairs = [(ancestor, node) for node in causal_graph.nodes for ancestor in nx.ancestors(causal_graph, node)]
if adjacent_only:
    pairs = [(ancestor, node) for (ancestor, node) in pairs
             if ancestor in get_ordered_predecessors(causal_graph, node)]

For a graph with N=100 nodes that is called repeatedly (e.g., during falsification), this meant 100 Python function calls and 100 separate graph traversals before any independence tests even started.

Changes

adjacent_only=True — skip ancestor enumeration entirely; direct edges are already available as causal_graph.edges():

if adjacent_only:
    pairs = list(causal_graph.edges())   # O(E) vs O(N·(V+E))

adjacent_only=False (default) — use nx.transitive_closure_dag(), which adds edge (u, v) iff u can reach v (i.e. u is an ancestor of v). A single optimised graph traversal replaces N separate nx.ancestors() calls:

else:
    tc = nx.transitive_closure_dag(causal_graph)
    pairs = list(tc.edges())

nx.transitive_closure_dag has been available since NetworkX 2.7, which is below the project's minimum of >=2.8.5.

Impact

Scenario	Before	After
100-node graph, adjacent_only=False	100 BFS/DFS calls	1 optimised traversal
100-node graph, adjacent_only=True	100 BFS/DFS calls + filter	1 edge iteration
Correctness	✓	✓ (proven: sorted(old) == sorted(new) on test graphs)

Test Status

Two new unit tests added in tests/gcm/test_falsify.py:

• test_given_chain_dag_when_validate_pd_without_adjacent_only_then_tests_all_ancestor_pairs — verifies N_TESTS=3 for a 3-node chain (all ancestor pairs)
• test_given_chain_dag_when_validate_pd_with_adjacent_only_then_tests_only_direct_edges — verifies N_TESTS=2 for a 3-node chain (only direct edges)

Full CI suite requires Poetry with all optional extras (not available in this environment). The optimisation is semantically equivalent: sorted(old_pairs) == sorted(new_pairs) was verified with a standalone NetworkX 3.4.2 script.

Generated by 🌈 Repo Assist, see workflow run. Learn more.

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/repo-assist.md@11c9a2c442e519ff2b427bf58679f5a525353f76