Title: Stationarity of Time-Series on Graph: A Generalized Approach via Transition Invariance

Abstract: Stationarity is a cornerstone in classical signal processing (CSP) for modeling and characterizing various stochastic signals for the ensuing analysis. However, in many complex real world scenarios, where the stochastic process lies over an irregular graph structure, the CSP is incapable of handling such signals. It is essential to establish a new framework to analyze the high-dimensional graph structured stochastic signals by taking the underlying structure into account. To this end, looking through the lens of operator theory, we first propose a new bivariate joint transition operator (JTO) consistent with the abstract form of translation operators in other signal domains. Moreover, we characterize time-vertex filtering based on the proposed JTO. Thereupon, we put forth a new definition of stationarity in time-vertex domain using the proposed JTO together with its spectral characterization. Then a new joint power spectral density (JPSD) estimator, called generalized Welch method (GWM), is presented. Simulation results are provided to show the efficacy of this JPSD estimator. Then, by modeling the brain Electroencephalography (EEG) signals as time-varying graph signals, we use JPSD as the feature for the challenging task of emotion recognition. Experimental results demonstrate that JPSD yields superior emotion recognition accuracy in comparison with the classical power spectral density (PSD) as the feature set. Eventually, we provide some concluding remarks.