parameter k in SAGEConv

[vivelakorea]

Hi,

I've been investigating ways to aggregate neighbors for SAGEConv with the goal of encompassing not only the closest immediate neighbors but also those that are k steps away.

Although the parameter 'k' in SAGEConv lacked a clear explanation, I found that by specifying 'k=...' in the code, it executed successfully.

Could you kindly provide an explanation for the 'k' parameter in below code?

 from torch_geometric.nn import Linear, SAGEConv
 import torch.nn.functional as F

 class GNN(torch.nn.Module):

    def __init__(self,n,k):
        super(GNN, self).__init__()

        self.sage1 = SAGEConv(38,n,k=k,improved=True)
        self.out = Linear(n,1)

    def forward(self, data):
        x, edge_index= data.x, data.edge_index

        x = F.elu(self.sage1(x, edge_index))
        x = self.out(x)

        return x

[akihironitta]

As far as I understand the codebase, the kwarg k will be simply ignored. Instead, I'd stack k layers in the model to aggregate nodes that are k steps away from each seed node.

Out of curiosity, where have you found the usage of SAGEConv(..., k=k)?

[vivelakorea]

Actually one of my colleague recommended it, so I have no clue..