[Repo Assist] perf(gcm): avoid pre-allocated DataFrame in compute_data_from_noise and compute_noise_from_data

[github-actions]

🤖 This is an automated pull request from Repo Assist, an AI assistant.

Summary

Extends the dict-of-arrays accumulation pattern (introduced in #1544 for fitting_sampling.py) to two more functions in dowhy/gcm/_noise.py.

Affected functions

compute_data_from_noise and compute_noise_from_data both used the pre-allocated DataFrame anti-pattern:

# Before (inefficient with pandas 2.x CoW)
data = pd.DataFrame(np.empty((noise_data.shape[0], len(sorted_nodes))), columns=sorted_nodes)
for node in sorted_nodes:
    ...
    data[node] = ...  # triggers copy on each assignment
return data

# After (no intermediate copies)
data: Dict[Any, np.ndarray] = {}
for node in sorted_nodes:
    ...
    data[node] = ...  # plain dict insertion
return pd.DataFrame(data, columns=sorted_nodes)  # single allocation

compute_data_from_noise — additional improvement

The parent-data lookup moved from DataFrame column-selection to direct dict access:

# Before: DataFrame slice → triggers another copy
data[get_ordered_predecessors(causal_model.graph, node)].to_numpy()

# After: build parent matrix directly from already-computed arrays
parents = get_ordered_predecessors(causal_model.graph, node)
np.column_stack([data[p] for p in parents])

np.column_stack on 1-D arrays produces the same (N, k) shape as the original .to_numpy() call, so the contract with .evaluate() is unchanged.

Why this matters

With pandas ≥ 2.0 copy-on-write semantics, assigning to a column of a pre-allocated DataFrame forces an internal copy of the backing block on every iteration. For graphs with many nodes, this adds O(N2) memory work. Accumulating results in a plain dict and constructing the DataFrame once at the end reduces that to O(N).

These two functions are called inside tight loops in get_noise_dependent_function → _get_exact_noise_dependent_function, and during compute_noise_from_data in counterfactual and noise-estimation workflows, so the saving compounds.

Trade-offs

• No API change — function signatures and return types are identical.
• np.column_stack with 1-D arrays produces identical shape to the original DataFrame .to_numpy().
• The noise_samples_of_ancestors function in the same file uses the same pre-allocation pattern but is intentionally left for a follow-up, as its partial-ancestor skip logic makes the change slightly more involved.

Test Status

Tests could not be run in this environment (no Python runtime with dependencies). CI will validate.

Relevant test files:

• tests/gcm/test_noise.py — directly exercises compute_data_from_noise and compute_noise_from_data
• tests/gcm/test_counterfactuals.py — exercises the counterfactual workflow that calls these functions
• tests/gcm/test_graph.py

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

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