Building Socially-Impactful Domain Knowledge Applications Using Graph Neural Networks

This paper commences by introducing two scientific applications of Graph Neural Networks (GNNs). Both applications are related to sustainability, which is a non-controversial subject matter. We will build our analysis toward developing more complex and socially-impactful applications that attract debates. Then, the paper highlights the unique capabilities of GNNs, as well as what they do not do. The challenges in applying GNNs to more complex and socially-impactful applications are discussed, along with 3 potential solutions: a) We discuss the concept of an Implied Graph, which can be used in modeling highly complex dynamics to increase our ability to solve socioeconomic problems. b) We show how Lagrangian Neural Networks (LNNs) can be applied to solve practical problems where there should be multi-level learning. c) The paper also explores the natural progression from the form of federated learning mentioned above, where cooperation is assumed, to future applications where cooperation cannot be taken for granted and where the centralized engine may need to serve certain “policing” functions.