A0741
Title: Tensor-variate spatial mixture models
Authors: Jeffrey Andrews - University of British Columbia Okanagan (Canada) [presenting]
Abstract: A framework that can account for spatial autocorrelation among a large number of predictors while using very few free parameters in the covariance structure of a Gaussian mixture model (GMM) is introduced. This is accomplished by assuming a linear spatial correlation/covariance structure: the coefficients of which can be estimated through generalized least squares. This model is expanded to allow for a factor analyzer structure in one of the directions of the tensor. This model is motivated through applications in medical physics and contrasts the approach with both the standard vector-valued GMM, as well as the matrix/tensor-variate GMM. An up-to-date accounting of ongoing work using this model and intended improvements are provided.
