Fitting graph models to big data

Many big data applications collect large numbers of time series. A first task in analyzing such data is to find a low- dimensional representation, a graph, which faithfully describes relations among the measured processes and through time. The processes are often affected by a relatively small number of unmeasured trends. This paper presents a computationally tractable algorithm for jointly estimating these trends and underlying weighted, directed graph structure from the collected data. The algorithm is demonstrated on simulated time series datasets.