Proper way to do Edge perturbation for node classification

[yorgancioglukaan]

Hi all,

I want to perform an edge perturbation alaysis where I keep removing an increasing portion of edges before perfroming node classification. (i.e delete 1/2 of edges from original graph, then 3/4, 7/8,...). I have 2 questions about this.

1. Is modifying RandomLinkSplit a good place to start, or is there a better way to do this? Keep in mind these splits will be completely seperate from train/val/test sets for the nodes.

2. I am concerned that with very high sparseness there will be unconnected components or even 0-degree nodes. How well does PyG handle this? Should I construct a minimum spanning tree and add those edges to the dataset just in case?

I would appreciate any guidance. Thanks!

[rusty1s]

I guess you want to use utils.dropout_edge for this. If this results in unconnected components/isolated nodes, then this is generally not a problem (except for that a node could only learn from itself or its component).

[yorgancioglukaan]

That makes sense. Thanks!