STenSr: Spatio-temporal tensor streams for anomaly detection and pattern discovery

The focus of this paper is anomaly detection and pattern discovery in spatio-temporal tensor streams. As an example, sensor networks comprising of multiple individual sensor streams generate spatio-temporal data, which can be captured in tensor streams. Anomaly detection in such data is considered challenging because of the potential complexity and high order of the tensor data from spatio-temporal sources such as sensor networks. In this paper, we propose an innovative approach for anomaly detection and pattern discovery in such tensor streams. We model the tensor stream itself as a single incremental tensor, for example representing the entire sensor network, instead of dealing with each individual tensor in the stream separately. Such a model provides a global view of the tensor stream and enables subsequent in-depth analysis of it. The proposed approach is designed for online analysis of tensor streams with fast runtime. We evaluate our approach for detecting anomalies under different conditions and for identifying complex data patterns. We also compare the proposed approach with the existing tensor stream analysis method (Sun et al. in ACM Trans Knowl Discov Data 2, 2008). Our evaluation uses synthetic data as well as real-world data showing the efficiency and effectiveness of the proposed approach.