Incremental causal network construction over event streams

This paper addresses modeling causal relationships over event streams where data are unbounded and hence incremental modeling is required. There is no existing work for incremental causal modeling over event streams. Our approach is based on Popper’s three conditions which are generally accepted for inferring causality – temporal precedence of cause over effect, dependency between cause and effect, and elimination of plausible alternatives. We meet these conditions by proposing a novel incremental causal network construction algorithm. This algorithm infers causality by learning the temporal precedence relationships using our own new incremental temporal network construction algorithm and the dependency by adopting a state of the art incremental Bayesian network construction algorithm called the Incremental Hill-Climbing Monte Carlo. Moreover, we provide a mechanism to infer only strong causality, which provides a way to eliminate weak alternatives. This research benefits causal analysis over event streams by providing a novel two layered causal network without the need for prior knowledge. Experiments using synthetic and real datasets demonstrate the efficacy of the proposed algorithm.