Learning Graphs with Multiple Temporal Resolutions

This paper proposes a framework for learning time-varying graphs with multiple temporal resolutions from multivariate time series signals. Our method estimates multiresolution graphs by a top-down approach: Graphs are learned from a segment of the time-series data corresponding to the desired temporal resolution, and we impose a constraint so that the learned graphs at the target temporal resolution are close to that in the lower temporal resolution. The proposed approach overcomes the problem of existing time-varying graph learning methods that must infer graphs in a single temporal resolution. Experimental results with synthetic data demonstrate that our method outperforms a baseline graph learning method.