can't reproduce the results in the paper

[haorannlp]

The reported results on Citeseer with VGAE is
AUC: 90.8(0.01), AP: 92.0(0.02).
But I can only get AUC:86.6, AP: 88.7.
on cora:
reported: 91.4 92.6
my results: 84.8, 87.8
Is this difference caused by the specific datasplits? Or other settings?

[stefanosantaris]

Me neither. I am using exactly the same tensorflow implementation referred to the readme and same parameters mentioned on the paper but I have no luck. I see similar results with you @haorannlp.

@tkipf should we modify something in the code in order to reproduce the results?

[tkipf]

Thanks for bringing this up. This is potentially related to this pull request: #25 — I currently don’t have capacity to test this myself but you could try rolling back this change and see if it affects results on Citeseer.

[stefanosantaris]

I reverted the code and now I can efficiently reproduce the results of the paper. Thanks @tkipf. I will try to fix this and create a PR so that it is available on master.

[stefanosantaris]

Problem solved and a fix can be found here #55

[haorannlp]

Thank you guys!