Window-limited procedures for sequential change-point detection

Music City Center 

Modern large-scale, high-dimensional, and complex streaming data call for computationally (and memory) efficient sequential change-point detection algorithms that are also statistically powerful when the pre- and post-change distributions are not specified. This gives rise to a computation versus statistical power trade-off, an aspect less emphasized in the past in classic literature. When the post-change distribution is unknown, the generalized likelihood ratio (GLR) statistic is commonly used. The seminal work by Lai (1998) considered the window-limited GLR statistic due to the practical memory constraint and established the optimal window length to be on the order of logarithm of the Average-Run-Length (ARL) performance. In this talk, I will present several new procedures for online change-point detection, with optimal choice of window length, that enjoy certain first-order asymptotic optimality. We demonstrate the use of the general framework on subspace change, network change detections, as well as neural network-based and Maximum Mean Divergence (MMD) based change-point detections.

High-dimensional change-point detection

Window-limited procedure