Different algorithms giving different results in `shortest_path`

It seems that, at least when masked arrays are passed to shortest_path and some entries are 0, the output is algorithm-dependent.

In my application, the meaning of 0 in a dense graph to be passed to shortest_path should be that of an edge with no cost, not one with an infinite cost (absent edge). I understand that this is not the view adopted by shortest_path, and have hence resorted to using masked arrays instead. However, this leads to unexpected discrepancies in the results.

Reproducing code example:

import numpy as np
from scipy.sparse.csgraph import shortest_path

csgraph = np.array(
    [[0, 1, 0], 
     [1, 0, 0],
     [0, 0, 0]]
)
csgraph_masked = np.ma.masked_invalid(csgraph)

shortest_FW = shortest_path(csgraph_masked, method='FW', directed=False)
shortest_J = shortest_path(csgraph_masked, method='J', directed=False)
shortest_D = shortest_path(csgraph_masked, method='D', directed=False)
shortest_BF = shortest_path(csgraph_masked, method='BF', directed=False)

Expected result:

The documentation does not suggest that, when the graphs are undirected and there are no negative values, outputs according to different methods should be different (and since the 'auto' option is available, as a user one would rather like a guarantee that this is not the case). However, in the above example, while the last three results are all equal to

array([[0., 0., 0.],
       [0., 0., 0.],
       [0., 0., 0.]])

(which, incidentally, is the one I want), shortest_FW equals

array([[ 0.,  1., inf],
       [ 1.,  0., inf],
       [inf, inf,  0.]])

What I guess is happening is that 'FW' is converting to a dense representation and forgetting about the mask completely.

Scipy/Numpy/Python version information:

1.4.1 1.18.4 sys.version_info(major=3, minor=8, micro=2, releaselevel='final', serial=0)

[jakevdp]

I think the issue is here: https://github.com/scipy/scipy/blob/655ce1bd180e2554a12d501fea0f24a0c6e8123a/scipy/sparse/csgraph/_shortest_path.pyx#L302-L304

This logic doesn’t account for masked arrays with explicit zeros.

[jakevdp]

forgetting about the mask completely.

I don’t think the mask is being ignored. All algorithms start by calling validate_graph, which handles masked inputs: https://github.com/scipy/scipy/blob/655ce1bd180e2554a12d501fea0f24a0c6e8123a/scipy/sparse/csgraph/_validation.py#L28-L34