Training and Evaluating Graph Generative Models

In this thesis-by-articles we make several contributions related to graph generative models (GGMs) and their applications.

In our first article, we investigate the use of GGMs for the sequential design of 3D structures. We propose LEGO as a toy problem that is complex enough to approximate real-world design and develop a method for representing simple LEGO structures as a graph to train a GGM. We extend a popular GGM approach, Deep Generative Models of Graphs (DGMG), to operate on these LEGO structures and propose several evaluation metrics originally developed for generative models of images that utilize a relevant pretrained network.

In our second article, we dive deeper into the proposed metrics to identify a single metric that can be used to evaluate GGMs regardless of domain. We propose replacing the pretrained network used in evaluation with a randomly initialized one to reduce overhead requirements, and design several experiments to objectively score each metric on several criteria. Through this process, we show that pretraining the network is not required: a network with random parameters is sufficient for evaluation. We further identify several strong metrics that can be used to easily evaluate GGMs across domains.