Why the classfication accuracy on test set oscillates？

[TranSirius]

Hi,

Thanks for your codes. I have run your codes for more than 100 times on Cora. Strangely, the accuracy on test set oscillates severely. The average accuracy cannot achieve the reported average accuracy (83.0%). I am wondering if you have filtered out those abormal records or I may miss some details?

I would appreciate it that you can offer any instructions.

There are minor changes to your code:

• before creating a session

• save the results to a file (I name it 'sigmoid' because I have written another copy of code in which I substitute softmax with sigmoid)

The recording file is : gat_sigmoid.txt (You may find some experiment with accuracy less than 50%)

Much thanks :)

Zijun

[TranSirius]

It seems that I missed the early stop mechanism, which will save the currently best model to the file. I have run several copies of codes simultaneously and the model saving file was covered.

Thanks for your codes again :)