Fast Online Changepoint Detection via Functional Pruning CUSUM Statistics
Authors: Gaetano Romano, Idris A. Eckley, Paul Fearnhead, Guillem Rigaill
JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show how FOCu S can be applied to a number of different changes in mean scenarios, and demonstrate its practical utility through its state-of-the-art performance at detecting anomalous behaviour in computer server data. ... In Section 5 we show a monitoring application for FOCu S on some AWS Cloudwatch server instances. ... In Figure 5 we report log-ratios of the average detection delay for pairs of methods against the different change magnitudes. |
| Researcher Affiliation | Academia | Gaetano Romano EMAIL Department of Mathematics and Statistics, Lancaster University, Lancaster, United Kingdom, LA1 4YF Idris A. Eckley EMAIL Department of Mathematics and Statistics, Lancaster University, Lancaster, United Kingdom, LA1 4YF Paul Fearnhead EMAIL Department of Mathematics and Statistics, Lancaster University, Lancaster, United Kingdom, LA1 4YF Guillem Rigaill EMAIL Universit e Paris-Saclay, CNRS, INRAE, Universit e d Evry, Institute of Plant Sciences Paris-Saclay (IPS2), 91405, Gif-sur-Yvette, France and Universit e d Evry val d Essonne, CNRS, INRAE, Laboratoire de Math ematiques et Mod elisation d Evry, 91000, Evry, France |
| Pseudocode | Yes | Algorithm 1: FOCu S0 (one iteration) ... Algorithm 2: Algorithm for max{0, Qn 1(µ) + µ(xn µ/2)} for µ > 0 |
| Open Source Code | Yes | Software implementing FOCu S and the code for our simulation study is available at https://github.com/gtromano/FOCu S. |
| Open Datasets | Yes | We now evaluate FOCu S by comparing with a bespoke anomaly detection algorithm on the Amazon CPU utilization datasets from the Numenta Anomaly Benchmark (Ahmad et al., 2017). |
| Dataset Splits | Yes | A method can use the first 15% of each dataset, a portion of data known to not include any anomalies, to set tuning parameters. |
| Hardware Specification | No | Runtimes were recorded for multiple finite sequences of lengths ranging from 100 to 5 104. To produce a fair comparison both implementations were written in C++, and all simulations were performed on a common personal computer. |
| Software Dependencies | No | To produce a fair comparison both implementations were written in C++ |
| Experiment Setup | Yes | For the Page recursion we employ a geometric grid (as recommended by Chen et al., 2022). We use a 10 point grid as that is equivalent to the expected number of intervals stored in FOCu S0 over a sequence of one hundred thousand observations. To see the potential benefits of using a finer grid, we also use a 20 point grid. ... A method can use the first 15% of each dataset, a portion of data known to not include any anomalies, to set tuning parameters. We use this data to tune both K in the biweight loss and the detection threshold as described in Appendix D. ... at each detection we inflate the threshold by a factor of log(τs)/ log(τs τs 1), with τ0, ..., τk being the estimated changepoint locations. |