Spatial deformation in a Bayesian spatiotemporal model for incomplete matrix-variate responses

2025-11-22

AI Summary

To address pervasive spatial anisotropy and nonstationarity in environmental and ecological data, this paper proposes a flexible matrix-variate Bayesian spatiotemporal model. Methodologically, it relaxes the conventional isotropy assumption via spatial deformation, integrates a matrix-variate covariance structure with a dynamic linear model, and performs joint spatiotemporal inference within a fully Bayesian framework; it further introduces a novel multivariate missing-data strategy that preserves full spatiotemporal dependence and enables efficient MCMC sampling. Experiments demonstrate substantial improvements in interpolation accuracy under strong anisotropy, robust performance under near-isotropic conditions, and scalability to large-scale incomplete observations. The core contribution lies in unifying the modeling of spatial heterogeneity, directional dependence, and multivariate missingness—thereby extending the capability of Bayesian analysis for high-dimensional spatiotemporal data.

Technology Category

• Machine Learning: Matrix & Tensor Methods
• Data Mining & Knowledge Management: Mining of Spatial, Temporal or Spatio-Temporal Data
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Graph Algorithms and Modeling for the Web: Algorithms and analysis for heterogeneous, signed, attributed, multi-relational, temporal, higher-order, and annotated Web-related graphs
• Web Mining and Content Analysis: Models for Web evolution
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

In this paper, we propose a flexible matrix-variate spatiotemporal model for analyzing multiple response variables observed at spatially distributed locations over time. Our approach relaxes the restrictive assumption of spatial isotropy, which is often unrealistic in environmental and ecological processes. We adopt a deformation-based method that allows the covariance structure to adapt to directional patterns and nonstationary behavior in space. Temporal dynamics are incorporated through dynamic linear models within a fully Bayesian framework, ensuring coherent uncertainty propagation and efficient state-space inference. Additionally, we introduce a strategy for handling missing observations across different variables, preserving the joint data structure without discarding entire time points or stations. Through a simulation study and an application to real-world air quality monitoring data, we demonstrate that incorporating spatial deformation substantially improves interpolation accuracy in anisotropic scenarios while maintaining competitive performance under near-isotropy. The proposed methodology provides a general and computationally tractable framework for multivariate spatiotemporal modeling with incomplete data.

Problem

• Modeling incomplete matrix-variate spatiotemporal data with flexible covariance structures
• Relaxing unrealistic spatial isotropy assumptions in environmental processes
• Handling missing observations while preserving joint data structure integrity

Innovation

• Spatial deformation method for nonstationary covariance structure
• Dynamic linear models in Bayesian framework for temporal dynamics
• Strategy handling missing observations preserving joint data structure