a bug in ismember function? about gae

[tkipf]

Thanks for catching this! This is indeed a bug and I have no idea what happened here (obviously a bug like this shouldn't happen with proper coding / unit testing practices).

The ismember function should look like this instead:

def ismember(a, b, tol=5):
    rows_close = np.all(np.round(a - b[:, None], tol) == 0, axis=-1)
    return np.any(rows_close)

I re-ran the Cora link prediction experiments with it and it looks like results are quite similar (I got slightly better results now than reported in the paper, but this could just be an outlier). I'll investigate this further. Feel free to make a pull request if you'd like to be listed as a contributor- otherwise I'll push the changes myself.

[tkipf]

Fixed with cc5aa6c