CCAD: A Collective Contextual Anomaly Detection Framework for KPI Data Stream

Ganghui Hu, Jing Wang, Yunxiao Liu, Wang Ke, Youfang Lin

Abstract: KPI (Key Performance Indicator) anomaly detection is critical for Internet companies to safeguard the availability and stability of services. Many online anomaly detection algorithms have been proposed in recent years. However, they are mainly designed to detect point anomalies, which fail to identify collective or contextual anomalies very well. Thus in this paper, we propose a framework of Collective Contextual Anomaly Detection (CCAD) for KPI data stream. Via Pearson correlation coefficient-based method to adapt to KPI stream, our framework addresses the limitation of time series discords algorithm and achieves a huge improvement in the accuracy for KPI stream. Moreover, instead of using a static threshold, we employ SPOT to generate an automatic threshold to determine anomalies. Extensive experimental results and analysis on multiple public KPIs show the competitive performance of CCAD and the significance of collective contextual anomalies.