DenseSAGEConv vs DenseGraphConv

[smorad]

I'm having trouble seeing how the implementation of DenseSAGEConv is different from DenseGraphConv. It seems like both do out = adj @ x, then push the nodes and neighbours through distinct FC layers lin_root(x) and lin_rel(out). Isn't SAGE supposed to do skip connections and generally be different than GraphConv?

[rusty1s]

Yes, they are both pretty similar. One difference is the normalize=True flag. Furthermore, DenseGraphConv has support for different aggregations (such as add), while DenseSAGEConv always performs (unweighted) mean-aggregation.