• graph generation
• recurrent neural network
• seq2seq
• variational attention
• variational autoencoder

Generating graphs is certainly a complex task. It requires to sample from a learned distribution of graphs. The current state-of-the-art is represented by an autoregressive model named GraphGen which approximates the distribution of graphs with a distribution of sequences.
To do so, it performs a canonization process to transform graphs into unique sequences which are then modelled through RNNs and MLPs.
This thesis proposes one modification and one extension to GraphGen. The modification is applied to the graph canonization phase to improve the data representation. The extension, instead, tackles the conditional graph generation task. This task is modelled as a seq2seq whose goal is to generate graphs similar to the input one. The proposed architecture is made of GraphGen used as the decoder of a variational autoencoder with variational attention.