Not being a graph neural network expert myself, I'm not sure how to evaluate this paper. Is this a big change? Would we expect more improvement? It seems like including edge and node information in a more "sophisticated" way should be better than separating them, but the gains aren't significantly better than other techniques.