enyandai / nrgnn Goto Github PK

Hi, Dai
There is some inconsistence in the paper and the code listed below.

1---It is writen that f_p and f_e will pretrain in the paper, but in the code it seems you just compute the feature cos simmilarities to get the potential edge set at the very beginning(and this is very essential. Without this step the performance greatlydecreases). I don't see any pretrain step.

2---The total loss in the paper are composed with L_E(reconstruction loss), L_p(the crossentropy loss of the pseudo label predictor on training set) and L_G(the crossentropy loss of the final classifier). It is writen that argmin L_G + αL_E + βL_P
On contrary, the line 133 in NRGNN.py, "total_loss = loss_gcn + loss_pred + self.args.alpha * rec_loss + self.args.beta * loss_add", the loss_add is not consistent with the lossL_p. Apparently there are four components in the code, and the loss_pred is the L_P in the paper. Is there any details about loss_add in paper that i missed?

Thanks

Random seeds have a great impact on the performance of the model. When I remove the fixed random seeds, the performance of the model decreases significantly.
For example, runs = 10
noise=uniform dataset=cora acc = 0.7743±0.0099,
noise=pair dataset=cora acc=0.7538±0.0136,
noise=uniform dataset=pubmed acc=0.7385±0.0115,
noise=uniform dataset=pubmed acc = 0.7850±0.0155.
Can you publish more experimental details and relevant codes of contrast models?