Spatiotemporal Smoothness-based Graph Learning Method for Sensor Networks

Graph learning often boils down to discovering the hidden structure of data that is characterized by a graph. Most of the previous works mainly focus on static data processing. However, the distributed sensor network gives rise to space- time data which exhibits spatiotemporal correlation and evolves smoothly over time. In this paper, we address the problem of learning graphs from space-time sensing data. Based on a dynamic model that takes into account both the spatial and temporal correlated property in temporal evolution, we propose a spatiotemporal smoothness-based graph learning method (GLSS), which novelly introduces the spatiotemporal smoothness to the field of space-time data analysis. By simultaneously recovering data and refining graphs under the spatiotemporal smoothness prior, the graph learning accuracy can be efficiently improved. Experiments on synthetic data and real-world data in weather sensor networks demonstrate that joint space-time analysis in proposed GLSS can bring benefits to graph learning and outperforms current the state-of-the-art methods.