Automated graph representation learning: a survey

Abstract

Research on Graph Representation Learning has seen many innovations in the recent years given the close resemblance of real-world data to graphs and its already wide usage. Graph data are considerably more difficult to analyse than traditional Euclidean data structures, which is the gap that Graph Representation Learning aims to cater to. However, automating Graph Representation Learning is a research area that is only starting to gain the focus it deserves. Most of the work that has been made available is not adequately documented or clear in their approach. It would be greatly beneficial for such work to be thoroughly analysed and evaluated under a common criterion. This paper introduces a novel automated Graph Representation Learning workflow that is utilized by an ongoing research project called AutoGRL which is built by the authors. It provides a standardized and modular production-ready platform for automated Graph Representation Learning with a level of autonomy and performance that has not been discovered. This paper also delivers an analysis of the existing work in the automated Graph Representation Learning domain, the benefits and limitations of the existing work, including comparisons of features along with new and unique research directions that are currently under-explored in the field.