I have been using a network to generate graphs. The architecture that I have been using is the following:

enter image description here

In this figure, $D_1$ is the signal generator and $D_2$ is the graph topology generator, which is a square, symmetric matrix which indicates which node is connected to which. In this network, $l$ shows linear layers, $a$ shows activation functions. Here we are using leaky relu activation function.

The problem that I am experiencing is that after training the network, my output is only a chain of nodes, meaning that only subdiagonal and superdiagonal elements have non-zero values and it is very rare to have other forms of graph. I was wondering if anyone has a suggestion for improving the output. Note that my training data is diverse and has every kind of graphs.



You must log in to answer this question.

Browse other questions tagged .