Enforcing Sparsity

[nassimaab]

First, thank you very much for this great project!

I want to generate realistic variants of Citeseer, so I tried to use your GVAE implementation. However, sampling adjacency matrices from the probability matrix computed by the GVAE leads to very dense graphs (~ 600.000 edges). Is this behavior expected? Or am I missing something? By inspecting the probability matrix, I found that more than 50.000 entries had a probability greater than 0.9.
Should I explicitly incorporate a term that enforces sparsity? Or maybe a sampling scheme that imposes constraints on the number of edges?

Best regards,