The International Conference for High Performance Computing, Networking, Storage and Analysis
Comparing Algorithms for Detecting Abrupt Change Points in Data.
Authors: Cody L. Buntain (University of Maryland), Christopher Natoli (University of Chicago), Miroslav Živković (University of Amsterdam)
Abstract: Detecting points in data where the underlying distribution changes is not a new task, but much of the existing literature assumes univariate and independent data, assumptions often violated in real data sets. This work addresses this gap in the literature by implementing a set of change point detection algorithms and a test harness for evaluating their performance and relative strengths and weaknesses in multi-variate data of varying dimension and temporal dependence. We then apply our implementations to real-world data taken from structural sensors placed on laboratory a bridge and two years of Bitcoin market data from the Mt. Gox exchange. Though more work is necessary to explore these real-world data sets more thoroughly, our results demonstrate circumstances in which an online, non-parametric algorithm does and does not perform as well as offline, parametric algorithms and provides an early foundation for future investigations.