Stochastic processes on graphs: learning representations and applications

Abstract

In this work, we are motivated by discriminating multivariate time-series with an underlying graph topology. Graph signal processing has developed various tools for the analysis of scalar signals on graphs. Here, we extend the existing techniques to design filters for multivariate time-series that have non-trivial spatiotemporal graph topologies. We show that such a filtering approach can discriminate signals that cannot otherwise be discriminated by competing approaches. Then, we consider how to identify spatiotemporal graph topology from signal observations. Specifically, we consider a generative model that yields a bilinear inverse problem with an observation-dependent left multiplication. We propose two algorithms for solving the inverse problem and provide probabilistic guarantees on recovery. We apply the technique to identify spatiotemporal graph components in electroencephalogram (EEG) recordings. The identified components are shown to discriminate between various cognitive task conditions in the data.