Thomas Kipf

This is a very good question -- this approach works as the model is typically not fitting the training data perfectly and hence has a certain inductive bias that allows...

Yes, it will probably work better, but most likely not by much (depending on how large your dataset is). At the same time, this will "leak" validation and test set...

I recommend using a dense matrix multiplication in the first layer if you have a somewhat dense feature vector. On Thu, May 30, 2019 at 12:38 AM Wangzhen wrote: >...

Sounds correct to me. Simply don’t use an activation function on the last layer (as this is handled separately). On Sat 1. Jun 2019 at 05:46 Wangzhen wrote: > @tkipf...

In principle, using a weighted graph as an input should work without modifications to the framework. In order to use a binary classification target for predicting links, you would presumably...

Yes, picking a threshold on the edge weights and selecting positive/negative samples based on this threshold is the easiest way to adapt your data to fit with the GAE framework....

Yes, I recommend using a sigmoid activation function and mean-squared error as loss function in this case. On Tue, May 21, 2019 at 12:41 PM Trevor Shim wrote: > Hello,...

This is a very good question. In my experience this type of setup can work very well. We are currently preparing a report where we describe this setup in more...

Yes, you are would sub-sample the negative class (non-edges) so that both classes are balanced in the binary cross-entropy loss. This is the same technique that is used in word2vec,...

Thanks! We're planning to release the code for these projects as well (probably some time after the NIPS review process).