medil.evaluate.sfd

medil.evaluate.sfd(true_biadj: numpy.typing.NDArray, predicted_biadj: numpy.typing.NDArray, to_return: str = 'raw') → int | float | tuple[int, float][source]

Structural Frobenius distance sums difference of latent parents.

For a binary biadjacency matrix B, consider U = B’B, where U_ij counts the number of parents nodes i and j have in common (so U_ii is the assignment number, and diag(U) is a sufficient statistic for the graph under the 1-pure-child assumption). sfd(B_1, B_2) is the sum of the differences between U_ij for B_1 and B_2 (without double-counting for U_ji).

Parameters:

  • true_biadj (ndarray) – Ground-truth biadjacency matrix.

  • predicted_biadj (ndarray) – Learned biadjacency matrix.

  • to_return (str, optional) – "raw" (default), "normalized", or "both".

Returns:

Raw sfd (int) if to_return='raw', normalized nsfd (float) if to_return='normalized', or (sfd, nsfd) if to_return='both'.

Return type:

int or float or tuple