andrey.metrics.shd

andrey.metrics.shd(estimated, true, *, endpoint_aware=False)

Structural Hamming distance between two graphs (lower is better; shd(g, g) == 0).

The default structural count charges one edit per unordered pair whose (mark_i, mark_j) differs, so an inserted, deleted, or reversed edge each costs 1. endpoint_aware=True sums the per-endpoint mark mismatches instead – the standard PAG distance, where a reversal or a deleted directed edge costs 2 and tail/arrow/circle are distinguished.

Return type:

int

Examples

>>> import numpy as np
>>> shd(np.array([[0, 1], [0, 0]]), np.array([[0, 0], [1, 0]]))  # 0->1 vs 1->0
1