B1671
Title: Time series anomaly detection with patterns and structural breaks identification: A constrained clustering approach
Authors: Carlo Drago - University of Rome Niccolo Cusano (Italy) [presenting]
Abstract: Time series anomaly detection is an approach useful to identifying deviations from a specific pattern over time. A relevant problem is to identify correctly the structure of the time series and the structural change which have to be considered as important by itself. We combine the analysis of the time series components and the structural change identification in order to characterize adequately the time series considered. At this point we consider an approach based on constrained clustering in order to detects the specific deviations of the time series from the identified patterns. In particular a validation of the clusters found over time can be useful to the aim to find the relevant anomalies which can occur on the data. Finally we will discuss on the sensitivity analysis which can be necessary to identify the anomalies which are not sensitive to the method used. In order to present the approach we will use both simulated time series and real data.
Journal