Graph Learning for Spatiotemporal Dynamic Signal

We address the problem of learning hidden graph structure from spatiotemporal signals which are prevalent in distributed sensor networks. Based on a space-time representation model that takes into account correlated properties in dynamic evolution, we formulate the graph learning problem as a regularized multi-convex optimization problem. A correlation-aware and differential smoothness-based graph learning method (CADS) is proposed, which simultaneously estimates the time correlation of each vertex and refines the graph under the differential smoothness prior. The proposed method promotes such smoothness property in each learning step, leading to an improvement of learning accuracy. Experiments on synthetic and real-world datasets demonstrate the effectiveness of the proposed CADS which outperforms the state-of-the-art graph learning methods in classification and prediction tasks.