Graph Generation with Recurrent and Graph Neural Networks

Graph generation has applications as diverse as drug discovery, materials design, and code completion. In this paper, we propose a novel auto-regressive graph generation model, where graph generation is viewed as a decision process. The proposed model combines the power of graph neural networks (GNNs) with generative modeling techniques, and incorporates both graph topology and node features, allowing the generation of graphs with desired properties. Extensive experiments on molecule datasets demonstrate the effectiveness of our approach, achieving high validity, diversity, and similarity to the target molecules.

I. Introduction

Graphs are powerful representations for modeling complex relationships and structures in various domains, ranging from social networks and biological networks, to recommendation systems and knowledge graphs. Graph generative models aim to capture the underlying patterns, connectivity, and features present in real-world graphs, and generate new graph instances that exhibit similar characteristics. The ability to generate realistic and diverse graphs is of great importance in understanding and analyzing real-world phenomena, and it has been applied to code completion [1], materials design [2], and drug discoveries [3].