Simulation-Free Structure Learning for Stochastic Dynamics

Abstract

We introduce a principled approach for jointly recovering the underlying network structure and dynamic response of a physical system. We show that our simulation-free method, NGM-SF2M, not only exhibits improved scaling relative to NGM on progressively larger linear systems, but also consistently retrieves a competitive recovery of the underlying network structure. Moreover, we show that incorporating interventional data yields improved performance for inferring network interactions. Our results indicate that while RF recovers marginally more accurate network structure, NGM-SF2M yields improved performance on the joint task – dynamical inference and structure learning. In future work, we aim to extend our framework to higher-dimensional systems, real-world settings, and integrate multi-modal data such as chromatin accessibility.