Maximum likelihood estimation of burst-merging kernels for bursty time series

Various time series in natural and social processes have been found to be bursty. Events in the time series rapidly occur within short time periods, forming bursts, which are alternated with long inactive periods. As the timescale defining bursts increases, individual events are sequentially merged to become small bursts and then bigger ones, eventually leading to the single burst containing all events. Such a merging pattern has been depicted by a tree that fully reveals the hierarchical structure of bursts, thus called a burst tree. The burst-tree structure can be simply characterized by a burst-merging kernel that dictates which bursts are merged together as the timescale increases. In this work, we develop the maximum likelihood estimation method of the burst-merging kernel from time series, which is successfully tested against the time series generated using several model kernels. We also apply our method to some empirical time series from various backgrounds. Our method provides a useful tool to precisely characterize the time series data, hence enabling to study their underlying mechanisms more accurately.