Automated Learning from Graph Structured Data

Machine learning is an important technique to learn from graph structured data (GSD). However, since it is knowledge- and labor-intensive to pursue good learning performance, humans are heavily involved in every aspect of learning from GSD. Examples are property prediction for molecular graphs, product recommendations from heterogeneous information networks, and logical queries from knowledge base. To make machine learning techniques easier to apply and reduce the demand for experienced human experts, automated machine learning (AutoML) has emerged as a hot topic with great interest. At the same time, a special type of deep network model has been recently proposed, called graph neural networks (GNNs), which capture the dependence of graphs via message passing between the nodes of graphs and allow learning from GSD in a more principled manner. Moreover, they allow learning from molecular dynamics and meteorological simulation, which further extend the classical scope of GSD.

However, again due to such diversities in GSD, there is no unitary GSD learning model that can perform consistently across different tasks and datasets. As a consequence, how to design adaptive methods to learn from GSD in a task-aware and data-specific manner is an important problem. Fortunately, there are emerging technical tools from machine learning communities that have the potential to solve the above problems. Examples are automated machine learning, neural architecture search, meta-learning, and learning to lean. These methods can all help generalize learning methods that can exhibit abilities to learn well across different datasets and tasks. Thus, by proposing this Research Topic, we hope to draw interest from both academia and industry, with the goal to push learning methods for GSD to the next level.

In this research topic, we are generally interested in the theory, methodology, new learning algorithms, and applications of GNNs on GSD. Specifically, the following topics are more preferable:

• Architecture design and theoretical understanding of GNNs

• Applying, extending, or customizing automated machine learning / learning to lean / meta-learning / neural architecture search methods to GNNs

• Applications or methodology building on GNNs, including (but not limited to) complex networks, heterogeneous information network, knowledge base

• Hyper-parameter tuning / bi-level optimization for GNNs / optimization techniques for GSD

• Exploration of the behavior of GNN in novel settings, including (but not limited to) noisy / adversary / heterophilous

We also acknowledge the contribution of Zhitao (Rex) Ying in preparing and coordinating this project.