Advantage of traditional embedding methods (e.g., metapath2vec) compared to GNN?

[bean]

Since DGL also supports traditional embedding methods (deepwalk, metapath2vec), what is the advantage of these methods compare to GNN?

[bean]

Maybe when topology information is important embedding methods are better.

[mufeili]

As you said, GNNs can be strong when you have rich input features.

[minjie]

Hi, the question is definitely a very open one. Prof. Peng Cui gave a quite insightful talk on this in KDD’20. See his talk slides here: https://www.calvinzang.com/file/kdd20/KDD_tutorial_part2_network_embedding_and_GCN.pdf