About “Test Error” problem
Sangs3112 opened this issue · 5 comments
Hi, I found that the error in SimGNN's paper is 0.001, but the error in your implementation results is 0.02, there is a ten-fold gap in the middle. I don't know why,I hope someone can give me an answer, thank you very much.
Yeah! I also found the difference between them. Then I want to use more data to train the model, but it seems useless.
And I also want to know how to get the similarity about the test graph, if there is only node and edge.
@zcyyyy24 I later found out this code may have some problem in settings and there is no one to maintain it anymore. Not only that, but I also found that the code in the other repositories was not set up according to the SimGNN paper. Then implemented a pytorch version of the code myself using SimGNN's official paper Settings, all using the paper Settings, including learning rate, batch_size, normalization method, four-layer MLP, GCNs_dim and so on. Finally, the mse of AIDS700nef dataset reaches about 0.0019, which is close to that of the paper. If you need this code, you can contact me and I can open source it on my github.
@zcyyyy24 I later found out this code may have some problem in settings and there is no one to maintain it anymore. Not only that, but I also found that the code in the other repositories was not set up according to the SimGNN paper. Then implemented a pytorch version of the code myself using SimGNN's official paper Settings, all using the paper Settings, including learning rate, batch_size, normalization method, four-layer MLP, GCNs_dim and so on. Finally, the mse of AIDS700nef dataset reaches about 0.0019, which is close to that of the paper. If you need this code, you can contact me and I can open source it on my github.
同学你好我们可以加个微信吗？GitHub上消息通知不太方便
是测试误差，具体结果不记得了。我记得没有官方论文中的结果好，不过差距很小，差距在1e-3之内，可能跟我没有微调有关，至少数量级是没有问题的，也完全遵循官方设置。但是代码的话我得等几天才能整理好公开，最近手上事情还比较多，临近期末了忙于考试。如果有需要可以加我微信联系