The documentation about edge orientation is inconsistent. In the Creating Message Passing Networks tutorial, the main expression says that e𝑖,𝑗 denotes (optional) edge features from node 𝑖 to node 𝑗., the attached expression also suggests it. However, in documentation to MessagePassing.message(), the documentation says Constructs messages from node 𝑗 to node 𝑖 (this is actually true).

I

📚 Documentation

We should demonstrate the power of GNN in inductive learning. That is, when a new node comes to a graph, we can apply GNN to do a good prediction.

Describe the bug

In StellarGraph, some algorithms are in demo notebooks or scripts, without being listed in our main documentation. This could mean that a user reading our docs thinks they're not supported in StellarGraph.

I noticed line 139 in "lib/utils.py" should be modified
from "model = gensim.models.Word2Vec(Sentences(self.documents), size)"
to "model = gensim.models.Word2Vec(sentences=Sentences(self.documents), size=size)"
I guess the parameter list of the Word2Vec constructor changed...

In LINE's code, there might be a minor error in negative sampling part (if I understand it correctly).
https://github.com/THUDM/cogdl/blob/a69a969020b8aa41cfcd8ac54511984bc5b32d62/cogdl/models/emb/line.py#L133-L137

If index j for negative samples start at 1, then the number of negative sample should be self.negative-1. For example, if you set self.negative=5, 0 is not the negative sample (sin

import torch
from torch_cluster import graclus_cluster

row = torch.tensor([0, 1, 1, 2])
col = torch.tensor([1, 0, 2, 1])
weight = torch.Tensor([1, 1, 1, 1]) # Optional edge weights.

cluster = graclus_cluster(row, col, weight)
print(cluster)
tensor([0, 0, 1])

The result seems to be 'tensor([0, 0, 2])' in my mac

I viewed the whole code and found that the code only use toy dummy data to train model. So I don't really understand how you use those data to train GCN model. Can you supply the code or instructions about how to use real-world data to train model?