Ordinary differential equations on graph networks

Recently various neural networks have been proposed for irregularly structured data such as graphs and manifolds. To our knowledge, all existing graph networks have discrete depth. Inspired by neural ordinary differential equation (NODE) for data in the Euclidean domain, we extend the idea of continuous-depth models to graph data, and propose graph ordinary differential equation (GODE). The derivative of hidden node states are parameterized with a graph neural network, and the output states are the solution to this ordinary differential equation. We demonstrate two end-to-end methods for efficient training of GODE: (1) indirect back-propagation with the adjoint method; (2) direct back-propagation through the ODE solver, which accurately computes the gradient. We demonstrate that direct backprop outperforms the adjoint method in experiments. We then introduce a family of bijective blocks, which enables $\mathcal{O}(1)$ memory consumption. We demonstrate that GODE can be easily adapted to different existing graph neural networks and improve accuracy. We validate the performance of GODE in both semi-supervised node classification tasks and graph classification tasks. Our GODE model achieves a continuous model in time, memory efficiency, accurate gradient estimation, and generalizability with different graph networks.