Detecting Statistically Significant Temporal Associations from Multiple Event Sequences

Abstract

In this paper, we aim to mine temporal associations in multiple event sequences. It is assumed that a set of event sequences has been collected from an application, where each event has an id and an occurrence time. Our work is motivated by the observation that in practice many associated events in multiple temporal sequences do not occur concurrently but sequentially. We proposed a two-phase method, called Multivariate Association Miner (MAM). In an empirical study, we apply MAM to two different application domains. Firstly, we use our method to detect multivariate motifs from multiple time series data. Existing approaches are all limited by assuming that the univariate elements of a multivariate motif occur completely or approximately synchronously. The experimental results on both synthetic and real data sets show that our method not only discovers synchronous motifs, but also finds non-synchronous multivariate motifs. Secondly, we apply MAM to mine frequent episodes from event streams. Current methods are all limited by requiring users to either provide possible lengths of frequent episodes or specify an inter-event time constraint for every pair of successive event types in an episode. The results on neuronal spike simulation data show that MAM automatically detects episodes with variable time delays.