How to implement GAT by using neighbor sampling and min-batch? In Min-batch GCN, we can subtract the self part easily, but in GAT we cannot directly subtract the self part. Is there any method can handle this case? Thank you!
As of dgl 0.3, we now support built-in neighbor sampling which allows the option of adding self loops in the sampled subgraphs. Check our tutorial on the topic.
Thanks for the reply!
A reminder: though you could try GAT with neighbor sampling, there is no theoretical guarantee saying it would work.
GCN with some sampling strategy is a consistent estimator of the loss function(e.g. FastGCN), but it’s not the case for GAT.
Thanks for the reply!