Unable to reproduce performance

[Shania99]

Hello,

I am currently trying to replicate the performance metrics published in your paper. I used the test.py script with the provided model_bpCL.pt weights on the test_graph.pt data. However, I encountered a significant discrepancy in the Fmax value; I obtained an Fmax of 0.379, which is quite different from the expected 0.595 as reported.

I have set up the environment exactly as specified in the environment.yml and followed the other setup instructions as given.

I have attached a screenshot of the test results for your reference. Could you please help me understand what might be causing this issue? Is there any step or configuration that I might be missing?

[ZhonghuiGu]

I'm sorry that the Fmax and AUPR provided by the test script confuse you. The standard evaluation script is "evaluation_matrics.py", which consider all the accept threshold (probability between 0.01-0.99), as well as the ancestor function filling.