A Unifying Method for Outlier and Change Detection from Data Streams

Detection of outliers and identification of change points in a data stream are two very exciting topics in the area of data mining. This paper explores the relationship between these two issues, and presents a unifying method for dealing with both of them. This approach is based on a probabilistic model of time series whose parameters are updated adaptively. The forward and backward prediction errors over a sliding window are used to represent the deviation extent of an outlier and the change degree of a change point. Unlike former approaches, the present one uses fuzzy partition method and fuzzy decision principle to alarm possible outliers and changes, which is more appropriate for online and interactive data mining from data streams. Simulation results confirm the effectiveness of the proposed method