Graph self-supervised learning and generative model for modeling graph structural property

Abstract

The structural characteristics of a graph, such as connectivity, are important factors that determine its properties. My thesis focuses on self-supervised learning methods and generative models to explicitly learn the structural characteristics of graphs.

Firstly, we propose a self-supervised learning method that can discretize the embedding space depending on the discrete graph structure. The proposed method enables the discrimination between valid and invalid graphs and explicitly learns the differences in graph connectivity. Through experiments, we demonstrated that even small differences can be distinguished clearly, unlike traditional contrastive learning techniques.

Secondly, we proposed a generative model that explicitly models the characteristics of graphs when generating the graphs. To this end, we derive a destination-driven generation process and proposed a new objective function that predicts the destination of the generation process. The proposed method allows the generation of graphs with inherent properties and demonstrated the generation of valid graphs.