LRZ CONVOLUTION: An Algorithm for Automatic Anomaly Detection in Time-series Data

Automatic anomaly detection is a hard but practically useful problem. With telemetry data sizes growing constantly, experts will rely increasingly on automation to bring anomalies to their attention. In this paper, anomaly transition points (called change points elsewhere), are determined using a novel application of a somewhat obscure statistical score called "z-score of mean difference". Use of this score yields a practical linear-time algorithm called LRZ Convolution with sound statistical underpinnings, and which does not require data normality. Each anomaly transition point is accompanied by a set of explanatory predicates that can form a good starting point for determining an anomaly's root causes. Careful experimental evaluation and performance in two independent domains show promising results. A preliminary comparison with a well-known machine learning algorithm called Support Vector Machines (SVM) yields a highly favorable outcome.