WSINDy method learns network dynamics from multiple initial conditions

A team led by Moyi Tian at the University of Colorado Boulder has applied Weak Form Sparse Identification of Nonlinear Dynamics (WSINDy) to learn effective models from network dynamics data. Their study, published on arXiv (2605.30432), focuses on social systems where individuals influence each other through online and offline interactions. The researchers generated data using a mean-field approximation of a stochastic interaction process on networks, then tested how accurately WSINDy could recover the underlying equations under varying noise levels. A key finding: when noise is high, using more initial condition trajectories improves accuracy, but the marginal gain diminishes rapidly—only a small number of additional trajectories is needed to capture most of the benefit.

Beyond noise handling, the work demonstrates that WSINDy can directly learn ordinary differential equation (ODE) models from averaged stochastic network data. This is particularly valuable when traditional mean-field approximations break down, as the discovered ODEs better match the data and provide deeper insight into the dynamics. The paper spans 24 pages with 14 figures and covers applications in dynamical systems, machine learning, and social network analysis. By sidestepping the need for explicit model forms, WSINDy offers a data-driven alternative for understanding complex social behaviors, potentially improving predictions of opinion spread, disease transmission, or other network-mediated processes.