deepgraphlearning / gmnn Goto Github PK

Hi,

It is good to see a great code for the paper. It is possible to release the link prediction code?

Hello, I found that if I import related data sets from DGL or other libraries, the results are not as expected. Has the data set been processed?

I would like to ask whether could we implement link classification using GMNN ? Cause i notice that we first should construct a dural graph and reformulate link classification as an object classification problem.

Hi Meng, when running the code I found that the calculation for accuracy seems to be reporting the maximum accuracy from the 100 epoches. I pasted the get_accuracy function from semisupervised/codes/train.py here:

def get_accuracy(results):
best_dev, acc_test = 0.0, 0.0
for d, t in results:
if d > best_dev:
best_dev, acc_test = d, t
return acc_test

However, in your paper you said that "The mean accuracy over 100 runs is reported in experiment" (Page 7), not the highest accuracy.

Could you explain this problem of which calculation method you actually used for the results reported in the paper? Thank you!

Hello guys,
I found that GMNN runs with the binarized features, but other baselines run with the real-valued features. I think that the experimental results are not fair and proper.

Also, I run GCN with the binarized features, and it also can achieve good results. Under this settings, the margins between GCN and GMNN is very small.

Besides, I run GMNN with real-valued features, and the performances drop a lot.