GLS Kernel Regression for Network-Structured Data

33 Pages Posted: 30 Aug 2021

See all articles by Edward Antonian

Edward Antonian

Heriot-Watt University

Gareth Peters

University of California Santa Barbara; University of California, Santa Barbara

Mike J. Chantler

Heriot-Watt University - Department of Computer Science

Hongxuan Yan

Chinese Academy of Sciences (CAS)

Date Written: August 9, 2021

Abstract

In this paper we consider the problem of predicting a signal y_t, defined across the N nodes of a fixed graph, over a series of $T$ regularly sampled time points. At each moment the graph signal is postulated to be a function of a set of global features x_t, which exist independently of the graph structure. We extend recent work in the field by designing an O(N^3 + T^3) kernel regression algorithm where the assumption that each observation is independent and identically distributed is relaxed. The result is a computationally tractable class of GLS models for graph regression. Specific attention is paid to the case where the graph signal is modelled as an AR(1) vector autoregressive process, whilst also exhibiting cross-correlation amongst nodes. We use the Laplace approximation to provide a lower bound for the marginal out-of-sample prediction uncertainty in a form that is tractable for large problems. Finally, the method is compared against other learning approaches on real data taken from a network of air quality monitoring stations across California. Here the goal is to predict the concentrations of various different pollutants 24 hours in advance given weather and environmental conditions. We find that kernel graph regression algorithms (both GLS and regular) out-perform competing techniques, whilst also providing a lower bound on the output uncertainty derived from the Laplace approximation.

Keywords: Graph Signal Processing, Regression, Kernel

Suggested Citation

Antonian, Edward and Peters, Gareth and Chantler, Michael John and Yan, Hongxuan, GLS Kernel Regression for Network-Structured Data (August 9, 2021). Available at SSRN: https://ssrn.com/abstract=3901694 or http://dx.doi.org/10.2139/ssrn.3901694

Edward Antonian (Contact Author)

Heriot-Watt University ( email )

Riccarton
Edinburgh EH14 4AS, Scotland EH14 1AS
United Kingdom

Gareth Peters

University of California Santa Barbara ( email )

Santa Barbara, CA 93106
United States

University of California, Santa Barbara ( email )

Michael John Chantler

Heriot-Watt University - Department of Computer Science

Edinburgh
United Kingdom

Hongxuan Yan

Chinese Academy of Sciences (CAS) ( email )

52 Sanlihe Rd.
Datun Road, Anwai
Beijing, Xicheng District 100864
China

Do you have a job opening that you would like to promote on SSRN?

Paper statistics

Downloads
165
Abstract Views
607
Rank
346,440
PlumX Metrics