EcoSta 2022: Start Registration
View Submission - EcoSta2022
A0619
Title: A scalable Gaussian process for large-scale periodic data Authors:  Qian Xiao - University of Georgia (United States) [presenting]
Abstract: The periodic Gaussian process (PGP) has been increasingly used in various contexts to model periodic data due to its good performance. Yet, it has such a high computational complexity of $O(n^3)$ ($n$ is the data length) that the application of PGP is often obstructed for processing large-scale periodic data, such as speech signals, vibration signals and periodic motions. To address this challenge, we proposed a circulant PGP (CPGP) model which can greatly accelerate the computations of both parameter estimations and model predictions. The proposed CPGP decomposes the full likelihood into the sum of two computationally scalable composite likelihoods, and its computational complexity is $O(p^2)$, even $O(p log(p))$ for some special cases, where $p$ is a candidate period of PGP which is much smaller than $n$. Numerical examples are included to show the scalability and the computational efficiency of the proposed CPGP compared to some state-of-the-art methods. Simulation and real case studies are discussed to further illustrate the superiority of the CPGP.