A0224
Title: Continual density ratio estimation for online time series with applications in change detection
Authors: Yidong Ouyang - The Chinese University of Hong Kong, Shenzhen (China) [presenting]
Liyan Xie - Georgia Institute of Technology (United States)
Abstract: Density ratio estimation plays a crucial role in data-driven decision-making. A general setting with non-stationary data sequences is considered. Inspired by the telescoping density-ratio estimation (TRE) that can improve the estimation of ratios between two highly dissimilar densities, a continual density-ratio estimation (CDRE) framework is developed to track the density ratio over time. Sliding windows are constructed, and the density ratio is estimated between two consecutive windows, which will be used to update the density ratio estimate gradually. CDRE enjoys both computational and memory efficiency. Furthermore, a novel detection algorithm, called CDRE-CuSum, is also proposed to apply the CDRE outcomes for online change-point detection. The recursive structure of CDRE-CuSum statistics makes it efficient for online implementation. Empirically, the CDRE is demonstrated to perform well in tracking the density ratio for non-stationary time series and in detecting abrupt changes in data distributions.
