Clustering the nodes of the Cora dataset

[OliverSummers]

I want to cluster the nodes of the Cora dataset based on the in-degree value of the nodes so that each cluster only contains the nodes with the same in-degree value. I know how to get the counts:

from torch_geometric.datasets import Planetoid
from torch_geometric.utils import degree

dataset = Planetoid('./data','CORA')
d = dataset[0]
n = d.num_nodes
indegree = degree(d.edge_index[1], n, dtype=torch.long)
counts = torch.bincount(indegree)

But since I don't access the index value of the nodes, I don't know how to place each node in which cluster?

[rusty1s]

How about using argsort() to re-sort your data so that nodes with the same node degree have a contiguous memory layout? For example, for re-permuting the node features, you can do:

perm = indegree.argsort()
data.x = data.x[perm]

For re-permuting edge indices, you can do:

data.edge_index = perm[data.edge_index]

[OliverSummers]

Thank you for your useful suggestion. But still, I am not sure about how to place each node in a cluster since I don't access the index value of the nodes, any suggestions?

[rusty1s]

I see. Do you want to build individual subgraphs out of nodes with the same degree? In case yes, you can use torch_geometric.utils.subgraph(indegree = x, ...) to achieve that, where x is replaced by an integer.

If you solely want to group all node features, you can split the node features via torch.split based on your bincount vector after permutation.

[OliverSummers]

Thank you very much.