This paper studies the mining of patterns for predicting critical events from observed ordered event data, where the observations can contain interleaving from non-predictor and other predictor event sequences. These are characteristics of many practical applications such as monitoring in power systems or telecommunication networks, as well as computational biology. For settings where system behaviors are affected by noise, a critical event can sometimes occur without its predictor executed prior to it, and we propose algorithm to recursively compute the frequency that a predictor candidate precedes the critical event. This we use for identifying a predictor, and study the performance of such a scheme. We also consider the noise-free settings, in which a critical event occurs only after the execution of its predictor, and propose an algorithm to recursively compute the set of maximal predictors for each critical event.
