COMPSTAT 2018: Start Registration
View Submission - COMPSTAT2018
A0189
Title: A linear time method for the detection of point and collective anomalies Authors:  Alexander Fisch - Lancaster University (United Kingdom) [presenting]
Idris Eckley - Lancaster University (United Kingdom)
Paul Fearnhead - Lancaster University (United Kingdom)
Abstract: Anomaly detection is of ever-increasing importance for many applications such as fault detection and fraud prevention. This is primarily due to abundance of sensors within contemporary systems and devices. Such sensors are capable of generating a large amount of data, necessitating computationally efficient methods for their analysis. To date, much of the statistical literature has been concerned with the detection of point anomalies, whilst the problem of detecting anomalous segments --often called collective anomalies-- has been relatively neglected. We will introduce work that seeks to address this gap by introducing a linear time algorithm based on a parametric epidemic change point model. We present an approach that, with provable guarantees, is able to differentiate between both point anomalies and anomalous segments. Our computationally efficient approach is shown to be better than current methods, and we demonstrate its usefulness on the challenging problem of detecting exoplanets using data from the Kepler telescope.