Identifying anomalous cellular network events is important for network management. Timely identification can aid mitigation and help improve customer satisfaction. The data considered in this work is a space-time point process where the points are reports sent in by customers in response to unsuccessful cellular activity such as dropped calls. Our focus is on identifying, for each time period, clusters of such reports that represent unusually high volume relative to similar data in past time periods. This spatial anomaly detection is done using modifications of the K-scan method developed by Loh. The K-scan method uses a Ki statistic to identify candidate points for anomalous clusters. These Ki values are related to the inhomogeneous K function of the point pattern. We implemented a p-value approximation based on an extreme value distribution to replace bootstrap to assess significance. This improves the speed of the algorithm. We also use an FDR approach to account for multiplicity in time. This FDR approach was also used to implement a multi-scale version of K-scan, where multiple distance values are used in the Ki statistic. Finally, we investigated using an ellipse instead of a circle to define the Ki statistic used in K-scan, comparing the two methods with a simulation study. Results from applying K-scan to the data are also presented.
All Science Journal Classification (ASJC) codes
- Information Systems
- Computer Science Applications
- Anomaly detection
- Cellular network
- False discovery rate
- Gumbel distribution