Abstract:Gaussian processes (GPs) are commonplace in spatial statistics. Although many non-stationary models have been developed, there is arguably a lack of flexibility compared to equipping each location with its own parameters. However, the latter suffers from intractable computation and can lead to overfitting. Taking the instantaneous stationarity idea, we construct a non-stationary GP with the stationarity parameter individually set at each location. Then we utilize the non-parametric mixture model to reduce the effective number of unique parameters. Different from a simple mixture of independent GPs, the mixture in stationarity allows the components to be spatial correlated, leading to improved prediction efficiency. Theoretical properties are examined and a linearly scalable algorithm is provided. The application is shown through several simulated scenarios as well as the massive spatiotemporally correlated temperature data.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is the lack of flexibility in existing non - stationary Gaussian processes (NS - GPs) when dealing with spatial data. Specifically, although existing non - stationary models can generate non - stationarity, their changes in stationarity between different regions are not flexible enough. This leads to the possibility of over - smoothing or underestimating prediction uncertainty when using these models for spatial prediction. These problems are particularly important for spatial prediction because over - estimating spatial autocorrelation will lead to over - smoothing and underestimating prediction uncertainty; conversely, underestimating spatial autocorrelation will lead to excessive variance and inefficiency of the interpolator. To solve these problems, the paper proposes a Mixed - Stationary Gaussian Process (MSGP) model. This model constructs a non - stationary Gaussian process by setting separate stationary parameters at each location and uses a non - parametric mixture model to reduce the number of effective parameters. Different from a simple independent Gaussian process mixture model, this mixed stationarity allows for spatial correlation between components, thereby improving prediction efficiency. Specifically, the paper solves the following two key problems: 1. **Parameter Estimation Problem**: Instantaneous or piecewise stationary models usually require a large number of parameters, resulting in high variance in parameter estimation. Unless the number of stationary regions is artificially limited, it is difficult to avoid this problem. To solve this problem, the paper uses non - parametric Bayesian methods (such as mixture models) to reduce the number of parameters, so that different regions can share the same parameters without being continuous. This information borrowing significantly reduces the number of required parameters. 2. **Prediction Problem**: A simple piecewise structure has no correlation between any two regions, which can only use part of the data during interpolation and is a serious problem. Although the boundary can be smoothed by model averaging or marginalization, the paper points out that the prediction efficiency of this method is not as good as that of the method directly including Gaussian covariance. The proposed non - stationary model can be regarded as a mixture model depending on Gaussian processes, with dense matrix Gaussian correlation for a given component assignment. This makes it possible to cluster spatial data while maintaining prediction efficiency. Through the above methods, the paper provides a more flexible and efficient non - stationary Gaussian process modeling method when dealing with large - scale spatio - temporal correlated data.

Mixed-Stationary Gaussian Process for Flexible Non-Stationary Modeling of Spatial Outcomes

Active data based Gaussian process models for nonlinear spatiotemporal systems

Markovian Gaussian Process: A Universal State-Space Representation for Stationary Temporal Gaussian Process

ProSpar-GP: scalable Gaussian process modeling with massive non-stationary datasets

Latent Gaussian Process Regression

Bayesian inference for high-dimensional nonstationary Gaussian processes

A nonstationary soft partitioned Gaussian process model via random spanning trees

A spatial Gaussian process method for hyperspectral remote sensing imagery classification

Infinite-Horizon Gaussian Processes

Non-stationary Gaussian Process Surrogates

Spatio-Temporal GP Model Learning for Intention-Driven Motions.

Spatiotemporal clustering using Gaussian processes embedded in a mixture model

Fitting Spatial-Temporal Data Via a Physics Regularized Multi-Output Grid Gaussian Process: Case Studies of a Bike-Sharing System

A Robust Approach to Gaussian Processes Implementation

Compactly-supported nonstationary kernels for computing exact Gaussian processes on big data

Nonparametric Bayesian Mixed-effect Model: a Sparse Gaussian Process Approach

Stationarity without mean reversion in improper Gaussian processes

Efficient spatio-temporal Gaussian regression via Kalman filtering

Nearest-Neighbor Mixture Models for Non-Gaussian Spatial Processes

Correcting missingness in passively-generated mobile data with Multi-Task Gaussian Processes

Computer emulation with non-stationary Gaussian processes