A0976
Title: Towards scalable Gaussian process regression with a moment-based estimation procedure
Authors: Savita Pareek - Auburn University (United States) [presenting]
Lionel Voirol - University of Geneva (Switzerland)
Roberto Molinari - Auburn University (United States)
Abstract: Gaussian processes (GPs) are a powerful Bayesian non-parametric framework for supervised learning, enabling probabilistic function modeling and uncertainty quantification. By leveraging kernel functions, GPs effectively capture the underlying relationships in the data. However, likelihood-based estimation procedures for GPs incur a high computational cost of $O(n^3)$, making them impractical for large-scale datasets. To address this problem, various approximations have been proposed in the literature, such as global and local approximations (Nystrm, subset selection, sparse kernels, etc.). Nevertheless, the estimation of these models is still limited by their significant computational cost in large-scale datasets containing millions or billions of data points. A novel moment-based estimation procedure is proposed based on the wavelet variance and the generalized method of wavelet moments (GMWM) estimator. This framework achieves a log-linear computational complexity of O(nlog n), where n is the number of training points, making it highly scalable. The method is compared with the state-of-the-art implementations, such as Rs GauPro, and comparable predictive performance is observed with significant gains in computational time. Additionally, the method is applied to diverse real-world applications, and it is compared with existing software packages. The results demonstrate its practical scalability, establishing it as a promising solution for large-scale data modeling.
