Generation of new graph from trained GVAE

[BerardinoB]

Dear @tkipf, I really appreciate your work and I would like to adapt your code to my own use cases. In particular, I would like to generate new graphs by sampling from a learned latent space, as usually done with images in Variational Autoencoder models. In such a case, indeed, new data (images) might be generated by sampling from a latent space which is constrained to be Normal distributed. However, in your implementation, it is not really clear to me if this could be done. As far as I understood, the reconstruction of the original adjacency matrix is performed by an inner product of the embedded input z_mean. This imply that in order to generate new graphs, I cannot sample from a standard Normal distribution since there would be no trained layers to be used. Do I understood correctly? Is there any other way to train your model in order to sample from a Normal distribution after training the model?

Thanks in advance for your precious help. Bests,

[haorannlp]

You can sample from a standard normal distribution directly after training to generate new graphs

[BerardinoB]

Dear Haorannlp, as far as i understood,the way in which I can use this model is the following:

1. emb = sess.run(model.z_mean, feed_dict=feed_dict)
2. adj_rec = np.dot(emb, emb.T)

Now, the problem is that after training the model, I would like to generate new instances without using any graph as input of the model. In order to do this I need to produce the "emb" and perform the dot product. Now, I cannot figure out how can I obtain this "emb" just by sampling from a strandard Normal distribution. In fact, from the code I see that:

self.z = self.z_mean + tf.random_normal([self.n_samples, FLAGS.hidden2]) * tf.exp(self.z_log_std)

is the part that should be distributed as normal after training. Am I understood correctly? If the answer is yes and I sample from a normal distribution, I will obtain "z" and not "z_mean". Is that right?

[haorannlp]

Sampling from a standard normal distribution : self.z = tf.random_normal([1, FLAGS.hidden2]). You don't need encoder to generate new graphs, only the decoder is enough.