Workshop: High Performance Computing for Environmental and Earth Sciences (HPC4EES)
Authors: Lutong Li (School of Mathematics and Statistics, Beijing Institute of Technology) and Yiping Hong (School of Mathematics and Statistics, Beijing Institute of Technology; Tangshan Research Institute, Beijing Institute of Technology)
Abstract: Multivariate spatial data exhibit both spatial heterogeneity and spatial contiguity. Classical exploratory factor analysis (EFA), which assumes a globally homogeneous structure, is ill suited to capture these features. This paper proposes Bayesian spatially clustered factor analysis (BSCFA), which embeds a Bayesian approach within the spatially clustered factor analysis (SCFA) framework. BSCFA retains the spatially adaptive grouping mechanism of SCFA and shifts it from point estimation to posterior inference conditional on the estimated partition, providing conditional uncertainty quantification for group specific factor models given the final partition. In addition, conjugate priors introduce Bayesian regularization, which stabilizes group specific parameter estimation and is particularly beneficial when individual spatial groups contain limited samples. To overcome the computational burden of running full MCMC chains at each outer iteration, we develop a suite of optimization strategies including state projection, warm-start Gibbs sampling, adaptive short chain lengths, and group wise parallel execution, which collectively yield an approximately 23.6-fold speedup. Simulation studies demonstrate that BSCFA outperforms both EFA and SCFA in terms of loading covariance estimation accuracy and model selection criterion. Applied to built environment data from 1,535 stations in the Tokyo metropolitan area, BSCFA identifies spatially contiguous clusters with clear functional differentiation. The factor structures reveal regional variations in urban functionality, achieving a BIC reduction of 25.9% compared to SCFA and 31.2% compared to EFA.
Back to High Performance Computing for Environmental and Earth Sciences (HPC4EES) Archive Listing
Back to Full Workshop Archive Listing