Open ghost opened 2 years ago
Hi there,
Many thanks for pointing this out. The reason for this issue is that we follow the code from another repo and we overlooked this issue in the original code (I would have to admit the code seems sound at first glance). I fixed this issue on my side and did a quick evaluation on GENN. Here are the updated results:
The other A*-based methods need more time to evaluate so I will update a fix once the results are ready.
Finally, permit me to explain some more details. This issue also existed in the repo that we followed in 2020:
And gratefully the maintainer of Extended-SimGNN has fixed this issue in a commit in Jul 2021 (https://github.com/gospodima/Extended-SimGNN/commit/d59318cda7c3cc34fafab4b784ce3292464c6703).
Thanks for your quick response and correction. It is a sad story for both of us because my model is based on part of your code. haha
Thanks again and sorry for the inconvenience 😥😥
According to the paper:
However, the code is to calculate precision at the lowest k elements. And ground_truth is normalized GED which means the similarity.
So the code actually calculates the least similar ones, and the precision is thus outrageously high due to the uneven dataset (like AIDS).
https://github.com/Thinklab-SJTU/GENN-Astar/blob/b9e97a67590314598679988ef0a07798753257ac/src/utils.py#L35-L42
https://github.com/Thinklab-SJTU/GENN-Astar/blob/b9e97a67590314598679988ef0a07798753257ac/src/astar_genn.py#L844-L845
https://github.com/Thinklab-SJTU/GENN-Astar/blob/b9e97a67590314598679988ef0a07798753257ac/src/astar_genn.py#L825-L826
https://github.com/Thinklab-SJTU/GENN-Astar/blob/b9e97a67590314598679988ef0a07798753257ac/src/astar_genn.py#L679