Gaussian process generalised additive models (GAMs) are flexible regression tools, but their standard Matérn covariance kernels have a subtle flaw: the function variance depends jointly on the effect magnitude and the length-scale. This means smooth and rough components with identical amplitudes contribute very different amounts of variation. When each component gets an independent prior on its amplitude, the total variance is overestimated, particularly as the number of dimensions grows.
R2-GP, proposed by Javier Enrique Aguilar and colleagues, addresses this by taking direct control over each component's expected contribution to the prior explained variance. The method rescales each component according to its length-scale and the observed inputs, then models the rescaled variance as a share of the function variance induced by a global R2 prior. This design lets the model scale gracefully to many components.
To handle large datasets, the authors extend R2-GP to Hilbert space approximation methods, which exploit the spectral decomposition of kernels for fast inference. In simulations and real-world case studies, R2-GP achieved superior generalisation and parameter recovery compared with independent variance priors.