Active Sequential Change-Point Analysis of Multi-Stream Data

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

This project aims to develop efficient methodologies and algorithms for actively learning from high-dimensional streaming data under the sampling or resource constraints. In many real-world applications, a system consists of many processes that can generate many data streams. At some unknown time, an unusual event could occur to the system, for example, a disease outbreak, a manufacturing defect, or a fraud signal, yielding a set of anomalous processes. Most systems however are operated under resource constraints that prevent the simultaneous use of all resources all the time. Thus, the decision maker must be responsible for actively choosing which processes are prioritized for observation. This will enhance their existing knowledge about the occurring event or anomalous processes while exploring new information and accounting for the penalty of the wrong declaration. The research would have broader impacts in a wide range of real-world applications such as biosurveillance, epidemiology, engineering, homeland security, and finance. The project will integrate research and education by infusing research findings into the curriculum and by training graduate students. This project seeks to make comprehensive progress on methodology, theory, and application of active sequential change-point analysis of multi-stream data under the sampling or resource constraints. The specific research aims are: (1) design efficient active change-point detection algorithms with false alarm guarantees, (2) develop an asymptotic theory to characterize statistical performances of the developed methods, (3) post-hoc analysis to apply false discovery rate methods to identify anomalous processes; and (4) applications in sepsis screening with online monitoring data from medical sensors in intensive care units to identify sepsis patients as quickly as possible while avoiding alarm fatigue. Results of the project are expected to significantly advance the state of the art in sequential analysis, change-point detection, multi-armed bandit problems, streaming data analysis, and large-scale inference. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.