In the actual library implementation is possibile to learn the graph structure (the adjacency matrix) from the data together with the weights? I don’t know if this type of task has a name (maybe graph?), however, is it already built? Until now, I have seen only tutorials to learn weight matrix W and nothing about the adjacency matrix A.
1 Like
I’m not sure what exactly you mean, but there’s generated model to generate molecules, which can be considered as a graph.
Including:
DGMG: https://github.com/dmlc/dgl/tree/master/examples/pytorch/dgmg
JTNN: https://github.com/dmlc/dgl/tree/master/examples/pytorch/jtnn
Graph2Graph: https://github.com/wengong-jin/iclr19-graph2graph
1 Like
Hi Volpino,
There is some very recent research on learning the graph structure along with the graph neural network parameters: https://github.com/lucfra/LDS-GNN.
However, this implementation from the authors is in Tensorflow and I’m not sure if anyone has done a DGL version. Could be a feature request? 
1 Like
Hi all! I understood that learning the graph structure corresponds to learn the adjacency matrix together with the weights and that is usually known as dynamic graphs… So, if the actual gdl implementation doesn’t provide this case, can I do a feature request?
thanks!
Hey @raspberryice and @Volpino, feel free to raise a feature request. We also highly encourage you guys to contribute and we will offer help if you met problems.
1 Like
I would definitely love this is a feature request too! Great idea! I was considering implementing a similar model a dgl generative model.