Repository navigation
About the name #21
Description
Thank you very much for your interesting paper and code. I have a concern regarding the naming. Instead of pooling, I think what you proposed is a fully connected layer for graph. Each output node is connected with all the input nodes. The number of output nodes for diffpool can be larger or smaller than the input. Since it is officially called a pooling layer, all the subsequently proposed pooling layers should be compared with diffpool, which is actually based on a fully connected mechanism (a fc layer). It is unfair and doesn't quite make sense.
Activity
[-]About the name and datasets[/-][+]About the name[/+]RexYing commentedon Jul 25, 2019
Hi,
In terms of computation power, only pooled layers have fully connected structure, but the input level gnn does not. We tried to address your concern by calling this soft pooling, which will not be as efficient as having deterministic pooling (although clustering of big graphs via min cut or spectral method is very expensive). If there is fairness concern, I suggest taking into account of the clustering time for deterministic methods, before comparing efficiency.
Regarding the number of output nodes after pooling, it's always smaller than input, as defined by the assignment ratio, which is 0.1-0.25 in most cases.
Compared to alternatives such as GraphUNet, the downsampling ratio (assign ratio) is also much lower.
But GraphUNet has deterministic clustering, if you only want to compare with deterministic ones.
Best,
Rex