Localization, Decomposition, and Dictionary Learning of Piecewise-Constant Signals on Graphs (1607.01100v2)

Abstract: Motivated by the need to extract meaning from large amounts of complex structured data, we consider three critical problems on graphs: localization, decomposition, and dictionary learning of piecewise-constant signals. These graph-based problems are related to many real-world applications, such as localizing stimulus in brain connectivity networks, and mining traffic events in city street networks, where the key issue is to find the supports of localized activated patterns. Counterparts of these problems in classical signal/image processing, such as impulse detection and foreground detection, have been studied over the past few decades. We use piecewise-constant graph signals to model localized patterns, where each piece indicates a localized pattern that exhibits homogeneous internal behavior and the number of pieces indicates the number of localized patterns. For such signals, we show that decomposition and dictionary learning are natural extensions of localization, the goal of which is not only to efficiently approximate graph signals, but also to accurately find supports of localized patterns. For each of the three problems, i.e., localization, decomposition, and dictionary learning, we propose a specific graph signal model, an optimization problem, and a computationally efficient solver. The proposed solvers directly find the supports of arbitrary localized activated patterns without tuning any thresholds. We then conduct an extensive empirical study to validate the proposed methods on both simulated and real data including the analysis of a large volume of spatio-temporal Manhattan urban data. The analysis validates the effectiveness of the approach and suggests that graph signal processing tools may aid in urban planning and traffic forecasting.