On Graph Learning with Neural Networks

Abstract

Graphs are ideal for modeling natural systems where relations may be intrinsic among data objects. With massive data available, learning graph models from data has become potentially feasible as well as necessary. Yet from the traditional machine learning perspective, learning structural topology of an unknown graphical model remains challenging. In particular, it is computationally intractable to learn graph topologies beyond a tree structure. Nevertheless, deep learning with neural networks, showing great potentials in visual imagery and other application domains, offers an alternative venue for effective machine learning on graphs. In this review, we discuss graph (structure) learning with deep neural networks. In particular, we examine graph neural networks (GNNs) from the task-based and the architecture-based perspectives, respectively.