A0711
Title: FOCuS: Online changepoint detection with a constant per-iteration computational cost
Authors: Gaetano Romano - Lancaster University (United Kingdom) [presenting]
Idris Eckley - Lancaster University (United Kingdom)
Paul Fearnhead - Lancaster University (United Kingdom)
Guillem Rigaill - Universite Evry (France)
Abstract: Changepoint analysis has been of major interest in recent times, with an increasing number of applications demanding an online analysis of a data stream. And as one enters the real-time domain, several challenges appear that render most of the current methods infeasible. We will present the FOCuS procedure, a fast online changepoint detection algorithm based on the simple Page-CUSUM sequential likelihood ratio test, and show how it is possible to solve the online changepoint detection problem sequentially through an efficient dynamic programming recursion. The FOCuS procedure outperforms current state-of-the-art algorithms both in terms of efficiency and statistical power. Furthermore, the procedure was extended to allow for more general scenarios, such as the pre-change mean being unknown, or adding robustness to outliers via robust loss functions. We demonstrate FOCuS on the Amazon CPU utilization datasets from the Numenta Anomaly Benchmark, where the aim is to monitor and detect anomalous behaviors in the CPU utilization of various Amazon Cloudwatch instances in real-time.
