Stream prediction using representative episode rulEs

Stream prediction based on episode rules of the form "whenever a series of antecedent event types occurs, another series of consequent event types appears eventually" has received intensive attention due to its broad applications such as reading sequence forecasting, stock trend analyzing, road traffic monitoring, and software fault preventing. Many previous works focus on the task of discovering a full set of episode rules or matching a single predefined episode rule, little emphasis has been attached to the systematic methodology of stream prediction. This paper fills the gap by constructing an efficient and effective episode predictor over an event stream which works on a three-step process of rule extracting, rule matching and result reporting. Aiming at this goal, we first pro- pose an algorithm Extractor to extract all representative episode rules based on frequent closed episodes and their generators, then we introduce an approach Matcher to simultaneously match multiple episode rules by finding the latest minimal and non-overlapping occurrences of their antecedents, and finally we devise a strategy Reporter to report each prediction result containing a prediction interval and a series of event types. Experiments on both synthetic and real-world datasets demonstrate that our methods are efficient and effective in the stream environment.