Bayesian Gaussian Mixture Modeling for Symmetric Matrix Variate Data

Abstract

Statistical inference on individual activity networks has been a historically difficult task due to the lack of available data at the appropriate granularity and the complexity of modeling individual mobility patterns. The recent availability of GPS data from individual devices, combined with highly detailed demographic information, suggests that one of these challenges can now be addressed. We introduce a new model which we call the Symmetric Matrix-Variate Normal Mixture Model (STRUCTURED) to estimate how demographic traits influence changes in human activity networks, using sociomatrices that capture the probabilistic spatial overlap between individuals over time. We exploit the commutativity constraint inherent in the symmetric matrix-variate normal distribution to parameterize the column precision matrix as a polynomial of the row precision matrix, reducing the effective parameter space by an order of magnitude. We develop two variants of STRUCTURED: STRUCTURED-FP, which estimates the full polynomial, and STRUCTURED-RJ, which uses reversible-jump MCMC to select a reduced-order parameterization. Simulation studies demonstrate that STRUCTURED-RJ outperforms existing methods in sparse-data regimes, whereas STRUCTURED-FP is preferred when sample sizes are large. We apply the model to GPS-derived sociomatrices of 293 individuals in King County, WA, finding that local crime environments and youth employment density are the dominant demographic factors explaining variation in weekly activity overlap patterns.