Possible to obtain edges chosen by graclus?

[cameronosmith]

Hi, is there a way I can obtain the indices of the edges that graclus chooses?

Also, is EdgePool basically graclus clustering but with predicting the edge weights first?

[rusty1s]

graclus does not choose edges, it will cluster nodes. The resulting adjacency matrix is then given by the union of edges of clustered nodes.

EdgePooling is different in that it will compute a score for each edge, and will contract edges according to that score.

[cameronosmith]

But doesn't graclus pick which nodes to cluster based on the largest edge weight possible in the node's available neighborhood? In that sense isn't it like EdgePool?

[rusty1s]

Not entirely. graclus still iterates over nodes randomly, and picks the neighbor with highest edge weight to form a cluster for each random node, while EdgePool will globally sort edge scores and visit them in order.

[cameronosmith]

Right that's what I figured. But like EdgePool does in multiplying the contracted nodes by their contracted edge's score, is there a way to collect the indices of the edges chosen by graclus? Or should I just make a symmetric edge score and collect nodes in the same cluster to recompute the edge score to enable the edge score gradient to flow through them?

[rusty1s]

I see. If you are interested in the "edges" chosen by graclus, you can directly look for this in the cluster vector, i.e., two nodes will have the same index if they are clustered together. These pairs of nodes will denote the edges chosen by graclus.