The Configs to Achieve The Best Performance about extended-simgnn

[gospodima]

Hi @canqin001,
sorry for the late response. I tried different configs but I wouldn't be sure about best performance since I was restricted in time and therefore the learning time was not long enough to achieve a maximum performance for each config.
The most important points that I can mention: for my experiments on LINUX and AIDS datasets the Graph Isomorphism Operator gave better performance than the GCN, for IMDB GCN was better. However, the overfitting effect for GIN was also higher and on a longer learning period probably some play with dropout will be required. Also, usage of histogram (pairwise node comparison) caused a significant overfitting effect, so usually I turned it off.
Regarding other parameters I can't say much unfortunately.