I See You: An Anomaly Detection Framework for System Alerts
Jonathan Dunne, Akash Bhargava · 2021
One of the main challenges in anomaly detection is the broad way in which they are defined. While this problem has been studied over the past ninety years, previous anomaly detection studies focus on a technique to solve a broad range of use-cases. Our motivation to revisit this problem is the rise of system alerting within production queuing systems. We consider the problem of anomaly detection as a probabilistic one. Using an enterprise dataset, we address the question as to whether existing anomaly detection techniques can detect subtle changes in alert inter-arrival times, and if not can a robust technique be used to detect queue anomalies. Our results show that two existing techniques perform poorly (F1 = 0.38), and our two parameter ranking algorithm and probabilistic detection method identifies anomalies with a higher degree of precision.