Time Series Anomaly Detection Based on Data Stream Clustering

Abstract

With the development of network-based software system, companies and operators pay more and more attention to the analysis of key performance indicators (KPI), such as network traffic and user browsing time. The challenges of anomaly detection of KPI data include the lack of labels and the unexpected changes of time series data distribution, named concept drift. Stream data clustering method is also used to deal with the above situation of stream data. An anomaly detection method based on stream data clustering is introduces in this paper to deal with the challenges above. This algorithm is called similarity based stream clustering (SBSC), which aims to deal with seasonal time series and concept drift. By using the density based stream data clustering framework based on similarity as a distance measure for time series, SBSC can deal with seasonal time series well and detect the concept drift phenomenon of time series. Compared with several common unsupervised anomaly detection algorithms, the algorithm is more efficient at identifying seasonal time series anomalies on real KPI data, and it is easier to detect concept drift and subsequence anomalies.