Clustering Large Networks of Parametric Dynamic Generative Models

Analysis, prediction and control of parametric generative models for stochastic processes arise in numerous applications, such as in biology, telecommunications, geography, seismology and finance. In many of these applications, it is desirable to obtain an aggregated behavior from an underlying network of stochastic interactions. This paper focuses on the simplification of parametric models describing multiple stochastic processes, by aggregating the processes that have similar input-output behaviors in an ensemble. We propose a clustering-based method, which is general in the sense that the similarity metric upon which the aggregation relies can accommodate processes characterized by a variety of generative models. To illustrate our aggregation framework, we investigate an example system comprised of a set of point process models for earthquakes. Simulations are presented.