Graph Neural Stochastic Differential Equations: A Graph-Based Approach to Modeling Stochastic Differential Equations

[Submitted on 23 Aug 2023]

Download a PDF of the paper titled “Graph Neural Stochastic Differential Equations” by Richard Bergna and 3 other authors

Download PDF

Abstract: We introduce a new model called Graph Neural Stochastic Differential Equations (Graph Neural SDEs), which improves upon Graph Neural Ordinary Differential Equations (Graph Neural ODEs) by incorporating randomness through Brownian motion. This addition allows for the evaluation of prediction uncertainty, an important aspect often overlooked in current models. We focus on the variant called “Latent Graph Neural SDE” and demonstrate its effectiveness. Through empirical studies, we show that Latent Graph Neural SDEs outperform conventional models like Graph Convolutional Networks and Graph Neural ODEs, particularly in confident prediction, making them superior in handling out-of-distribution detection in both static and spatio-temporal contexts.