Finding periodic outliers over a monogenetic event stream

Abstract

Sensors are active everywhere. Enormous volumes of sensed events are sent over the data streams, while most of applications want to focus on events that would be curious. We propose a technique for mining periodicities and predicting its outliers from the stream. The key to our technique is a simple periodic pattern /spl Delta/t, derived from delta-time mining, or SUP(t, t+/spl Delta/t). We provide efficient algorithms for finding the highest support /spl Delta/t on a small and resource-limited sensor device. Our experiments compare memory efficiency and accuracy, on a variety of event patterns, monogenesis, polygenesis, and semi-random.