Gaussian Processes on Graphs Via Spectral Kernel Learning

We propose a graph spectrum-based Gaussian process for prediction of signals defined on nodes the graph. The model is designed to capture various signal structures through highly adaptive kernel that incorporates flexible polynomial function in spectral domain. Unlike most existing approaches, we learn such kernel, where setup enables learning without need eigen-decomposition Laplacian. In addition, this has interpretability filtering achieved by bespoke maximum likelihood algorithm enforces positivity spectrum. demonstrate synthetic experiments from which show ground truth filters can be accurately recovered, and adaptability translates superior performances real-world data characteristics.