Inferring a sequence of evolving graph topologies (weighted Laplacians) from multivariate signals whose distribution is heavy-tailed, so outliers do not corrupt the learned structure. The signal on the latent graph is modeled with a Student-t distribution for robustness, while the graph's temporal variation is driven by a non-negative vector auto-regressive (VAR) model; spectral constraints on the Laplacian promote clustered structure, and the estimator tolerates noise and missing values. Solved by an iterative stochastic optimization in a semi-online mini-batch framework, with applications to streaming financial data.
This paper learns time-varying graph topologies from data whose signals follow heavy-tailed distributions, a regime (e.g. financial time series) where standard Gaussian-based graph learning is badly …