Compare to the GCN embeddings

[alexanderpanchenko]

About

• https://tkipf.github.io/graph-convolutional-networks/

• https://towardsdatascience.com/how-to-do-deep-learning-on-graphs-with-graph-convolutional-networks-7d2250723780

Implementations

• pytorch: https://github.com/tkipf/pygcn

• TF: https://github.com/tkipf/gcn

• Keras: https://github.com/tkipf/keras-gcn

[alexanderpanchenko]

won't fix this this reason:

On Nov 15, 2018, at 17:26, Oleksiy Oliynyk oleksiy.oliynyk@gmail.com wrote:

regarding GCN:

I read the articles and this approach requires features of the nodes.

So in the implementation mentioned in the issue, they use 2 files as a input. https://github.com/tkipf/keras-gcn/tree/master/kegra/data/cora

one just contains pair of ids per line, but second has a) id --- it's obvious b) feature vector per every node (with binary values if feature is present or not) --- not sure what I can put here c) class --- we have just one class which is synset

Right, technically this approach in fact can’t be compared directly to our method as we do not really on the node features! In this case, I suggest to for now suspend this task.