The focus of this talk is on finite-state hidden Markov models (HMMs) where
either the transition probability matrix parameters or the observation
densities
(or both) change at a random changepoint. In the Bayesian setting, efficient
changepoint detection strategies are proposed to minimize the average
detection
delay subject to false alarm constraints. The intuition developed from the
theoretical study is leveraged in two application settings: detection of
sudden
spurts and downfalls in the activity profile of terrorist groups, and
detection
of changes in interaction dynamics in social media networks. Detection
strategies that exploit the underlying HMM structure are shown to
significantly
outperform naive strategies that do not exploit this structure.