Sequential Change-Point Detection for High-Dimensional and Non-Euclidean Data

In many applications, it is often of practical and scientific interest to detect anomaly events in a streaming sequence high-dimensional or non-Euclidean observations. We study non-parametric framework that utilizes nearest neighbor information among the observations changes an online setting. It can be applied data arbitrary dimension as long similarity measure on sample space defined. consider new test statistics under this more effectively than existing while keeping false discovery rate controlled at fixed level. Analytic formulas approximating average run lengths approaches are derived make them fast applicable modern datasets. Simulation studies provided support theoretical results. The proposed approach illustrated with analysis NYC taxi dataset.