Efficient Event Series Data Modeling via First-Order Constrained Optimization

Niccolò Dalmasso, Renbo Zhao, Mohsen Ghassemi, Vamsi K. Potluru, Tucker Balch, Manuela Veloso

Abstract: Event series data are ubiquitous in many domains, including finance, epidemiology, and advertising. Hawkes processes have recently emerged as prominent tools for both modeling and generating event series data. Specifically, multidimensional Hawkes processes model both self-excitation and cross-excitation between different types of events. In this work, we propose to learn multidimensional Hawkes processes using an adaptation of the Frank-Wolfe algorithm, which is a first-order constrained optimization method. Our approach is particularly suitable in sparse settings, i.e., when each event type only influences a small number of other event types. Empirical results on both simulated and real datasets show that our approach achieves better or comparable accuracy in terms of parameter estimation compared to other first-order methods and the state of the art for learning multidimensional Hawkes processes, while enjoying a significantly faster runtime.