Kriging via variably scaled kernels
New method breaks the stationary kernel assumption, enabling Gaussian processes to model abrupt changes and discontinuities.
A team of researchers has published a paper titled 'Kriging via variably scaled kernels' on arXiv, proposing a significant advancement for Gaussian process models. The work, led by Gianluca Audone, Francesco Marchetti, Emma Perracchione, and Milvia Rossini, tackles a core limitation of classical Kriging and Gaussian processes: their reliance on stationary kernels. These traditional kernels assume correlations between data points depend solely on their relative distance, which ensures analytical simplicity but fails to capture complex, heterogeneous patterns or abrupt changes in real-world data.
The new method introduces 'variably scaled kernels' as a tool to construct non-stationary Gaussian processes. By applying a scaling function, the technique explicitly alters the correlation structure between data points. This enables the modeling of phenomena with discontinuities or rapidly varying behavior that standard models struggle with. The authors analyze the resulting predictive uncertainty using a derived 'variably scaled kernel power function' and clarify the relationship between their construction and classical non-stationary kernels.
Numerical experiments presented in the paper demonstrate the practical benefits of this approach. Models built with variably scaled kernels achieve improved reconstruction accuracy compared to traditional methods. Crucially, they also provide uncertainty estimates that are more reflective of the true, underlying structure of the data, offering users a more reliable measure of confidence in predictions for complex spatial or temporal problems.
- Breaks the 'stationary kernel' assumption, allowing Gaussian processes to model heterogeneous correlation structures and abrupt changes.
- Uses a scaling function to explicitly modify data correlations, enabling the handling of targets with discontinuities.
- Numerical experiments confirm improved reconstruction accuracy and uncertainty estimates that better reflect the true data structure.
Why It Matters
Enables more accurate and reliable spatial modeling for complex real-world data in fields like geostatistics, environmental science, and engineering.