Thomas Kipf

I recommend having a look at "A Tutorial on Energy-Based Learning" by LeCun et al.. Using a scoring function like `exp(-d(x,y))` essentially boils down to an energy-based model with `d(x,y)`...

Yes, we've played around with model variants where we also reconstruct the node features (X) or include label information in a semi-supervised auto-encoder setting. Both approaches usually help improve the...

You can use a node-level MLP to reconstruct node features from the final decoder GCN outputs. If node features are binary, then simply using a binary cross entropy loss works...

PyTorch 0.4.0 is currently not supported (see README). Pytorch 0.2.0 is the officially supported version, however 0.3.0 might work as well.

Thanks, I appreciate your comment. Unfortunately, I won’t have time/capacity to work on the data loading pipeline. If you find a way to simplify the data loader, feel free to...

I recommend using the data loader from https://github.com/tkipf/keras-gcn/blob/master/kegra/utils.py Then you don’t have to deal with this strange .allx etc format (which is just supplied in this repo because it was...

Looks like your `a` in `a - b[:, None]` is of shape `(0,)`. Maybe step through it using a debugger and see where it goes wrong (i.e. where`a` gets this...

GAE is an inductive model: the decoder takes pairs of node embeddings to predict links, whereas the encoder dynamically constructs these node embeddings from node features and their local neighborhoods...

Batch-wise training should be possible using the approach described in this thread: https://github.com/tkipf/gcn/issues/4 -- you might have to explicitly exclude edges between different graphs in the decoder depending on what...

That seems to be a python issue unrelated to this repository. Not sure if/how I can help? On Wed 3. Oct 2018 at 16:03 Chao Jiang wrote: > I am...