The code of RGCN

[wzfhaha]

Dear authors,
I have noticed that the RGCN has an accuracy of 50.0% on MAG in your paper, which is higher than RGCN's result reported in OGB leaderboard. Could you share your codes for the RGCN model you used in the paper? Thanks very much!

[lingfanyu]

Hi @wzfhaha

I used the sampling version of RGCN(https://github.com/snap-stanford/ogb/blob/master/examples/nodeproppred/mag/sampler.py) from leaderboard.

The main change is we also generated features for those featureless nodes (like author, institution, topic), and increased the number of layers to match NARS. I think that's the main reason for the better accuracy.

The code for loading features into the graph is:

with torch.no_grad():
    rand_weight = torch.Tensor(128, 256).uniform_(-0.5, 0.5)
    data.x_dict["paper"] = torch.matmul(data.x_dict["paper"], rand_weight)
    data.x_dict["author"] = torch.load(os.path.join(args.emb, "author.pt")).float()
    data.x_dict["institution"] = torch.load(os.path.join(args.emb, "institution.pt")).float()
    data.x_dict["field_of_study"] = torch.load(os.path.join(args.emb, "topic.pt")).float()