Applications of Graph-Theoretic Tests to Online Change Detection
Trident Scholar Project rept. no. 424
NAVAL ACADEMY ANNAPOLIS MD
Pagination or Media Count:
Detecting change in a stochastic process is a central problem in statistics. This project explores nonparametric graph-theoretic approaches to solving online change-point problems. The foundation for our methodology is the Ensemble Sum of Pair-Maxima ESPM Test, a powerful offline test developed by Ruth and Koyak 2011. Our work investigates the efficacy of the ESPM Test in a variety of offline settings, and ultimately extends that test to online settings through a novel modification of recently developed multiple testing procedures designed to control false discovery rate. When tested against simulated and pseudo real-world data, this modified procedure maintains the desired overall test level while achieving impressive power and useful advanced warning times in many scenarios. This method is not limited to the ESPM test and holds much promise for adapting other powerful offline techniques to online scenarios.
- Statistics and Probability