A0863
Title: Joint learning of panel VAR models with low rank and sparse structure
Authors: George Michailidis - University of California, Los Angeles (United States) [presenting]
Abstract: Panel vector auto-regressive (VAR) models are effective in modeling the evolution of multivariate time series (with an identical set of variables) across different sub-populations. The aim is to develop a panel VAR model with shared low-rank structure modulated by sub-population specific weights, enhanced by idiosyncratic sparse components. Parameter identifiability issues are addressed through constraints that lead to a nonsmooth, nonconvex optimization problem. A multiblock ADMM algorithm is developed for parameter estimation and its convergence properties established under mild regularity conditions. Further, consistency properties under high dimensional scaling are also established for the parameter estimates. The performance of the posited model is evaluated both on synthetic data and on a neuroscience data set.
