Unifying Change -- Towards a Framework for Detecting the Unexpected

An interesting challenge in data stream mining is the detection of events where events are generally defined as anything previously unknown in the data. Therefore outliers, but also model changes or drifts, can be considered as possible events. Various methods for event detection have been proposed for different types of events. In this paper, we describe a more general framework for event detection. The framework enables generic types of time slots and streaming progress through time to be incorporated. It allows measures of similarity to included between those slots, either based directly on the data, or an abstraction, e.g. a model built on the data. We demonstrate that a large number of existing algorithms fit nicely into this framework by choosing appropriate time slots, progress mechanisms, and similarity functions.