zhuty16 / cigcn Goto Github PK

preprocess.py该文件debug后，存在一些问题🙋

To my understanding, I use the CIGCN's aggregation step as follow:
for hop in range(3):
agg_embed = F.tanh(torch.mul(torch.sparse.mm(D{-0.5}AD{-0.5}, X) , self.w_list[hop]))
And lightGCN's aggregation step is implemented as follow:
for hop in range(3):
agg_embed = torch.sparse.mm(D{-0.5}AD{-0.5}, X)
I think this is the only change between CIGCN and LightGCN.
I have read your paper and source code carefully, but the experiment results don't show any improvement but suffer degeneration. I only use user/item interactions without item/item interactions. But I don't think it's the reason causing such a degeneration under your paper's hypothesis.
Could you please answer my question and point out what I miss for your idea?

how to gen the .npy file

i get that .npy file from the internet ,but when i run the code ,it raise a problem that it needs a .csv file ....