Univariate Time Series Anomaly Labelling Algorithm

Abstract

Unsupervised anomaly detection in an n-length univariate time series often comes with high risk. Anomaly contextual dependencies limit the application of binary classification methods. Analyzing the statistical features of data may help enrich the context of anomaly detection. This article proposes a quadratic time algorithm for analyzing possible anomalies in the context of unsupervised learning. Detection of possible anomalies uses Median Absolute Deviation on the residual of a univariate time series. Computation of residuals uses robust STL (Seasonal and Trend decomposition using Loess). Experiments on three datasets (Yahoo, NUMENTA NAB and district-heating substation power profiles) show the ability of the algorithm to enrich anomalies by associating labels such as Certainty, Uncertainty, and Probable, with the probable class indicating a need to further process the anomalies.