How to sort Planetoid dataset based on the degree values of nodes?

[OliverSummers]

The Cora dataset has the following attributes:

Data(edge_index=[2, 10556], test_mask=[2708], train_mask=[2708], val_mask=[2708], x=[2708, 1433], y=[2708])

What I want to do is to sort the dataset based on the degree value of nodes. I code as follows:

def sort_dataset(data):
   
    n = data.num_nodes
   
    degrees = degree(data.edge_index[1], n, dtype=torch.long)
    sorted, indices = torch.sort(degrees)
    new_x = data.x.detach().clone()
    new_y = data.y.detach().clone()

    for i in range(len(new_x)):
        new_x[i] = data.x[indices[i]]
        new_y[i] = data.y[indices[i]]

    data.x = new_x
    data.y = new_y

    return data

dataset.transform = sort_dataset(data)

One thing missing in my code is rearranging data.edge_index as well. So based on the body of "sort_dataset" I shared above, how do I rearrange data.edge_index?

Thank you.

[rusty1s]

deg = degree(edge_index[1], num_nodes, dtype=torch.long)
perm = deg.argsort()
x = x[perm]
edge_index = perm[edge_index]

should give you the desired result.

[OliverSummers]

Thank you. To get the new x and y, your code suggestion is much shorter than mine but I think there is still an issue with the new edge_index. When I sorted the dataset based on your suggestion and fed it into my GCN model, the accuracy was degraded to 23% from 81%. What is the issue in your opinion?

[rusty1s]

You are right. The above code has some bugs. This should fix it:

inv_perm = torch.empty_like(perm)
inv_perm[perm] = torch.arange(perm.size(0))

x = x[perm]
edge_index = inv_perm[edge_index]

[OliverSummers]

Thank you very much. Your solutions are amazing. One more question, if I want to sort the dataset based on the page ranking of the nodes or any other requirement, how should we do it? Here we use degree(edge_index[1], num_nodes, dtype=torch.long) as our requirement is the degree value of nodes, but in other requirement cases, I don't know what to use instead of PyTorch degree function.

[rusty1s]

If you have the PageRank values of every node in a vector, then sorting based on these values is similar to sorting based on deg (via argsort()). We currently do not provide functionality to compute PageRank values though. As an alternative, you can convert your data to networkx and compute PageRank values from there.

[OliverSummers]

Thank you.