Abstract
In this paper, we propose Graph Differential Equation Network (GDeNet), anapproach that harnesses the expressive power of solutions to PDEs on a graph toobtain continuous node- and graph-level representations for various downstreamtasks. We derive theoretical results connecting the dynamics of heat and waveequations to the spectral properties of the graph and to the behavior ofcontinuous-time random walks on graphs. We demonstrate experimentally thatthese dynamics are able to capture salient aspects of graph geometry andtopology by recovering generating parameters of random graphs, Ricci curvature,and persistent homology. Furthermore, we demonstrate the superior performanceof GDeNet on real-world datasets including citation graphs, drug-likemolecules, and proteins.