ÄûÃʵ¼º½

The title of the thesis: Detecting changes in distributions in large-scale streaming data

Thesis defender: Topi Halme
Opponents: Prof. Urbashi Mitra, University of Southern California, US, and Prof. Pierluigi Salvo Rossi, NTNU, Norway
Custos: Prof. Visa Koivunen, Aalto University School of Electrical Engineering

The rapid detection of sudden changes and anomalies in real-time data streams is a central challenge in many application domains. For example, in order for communication systems, environmental monitoring, or power grids to operate as efficiently as possible, sudden statistical changes and disturbances must be identified from noisy observations as quickly and reliably as possible. 

This doctoral dissertation addresses situations where multiple data streams are monitored in parallel, such as sensor measurements from different geographical locations or frequency bands. Depending on the application, a change may occur in one or several streams, either simultaneously or at different times in different places. The problem of real-time change-detection has been widely studied in the literature, but less so in high-dimensional, multi-stream settings, which present their own particular challenges. 

In change detection problems, not only is the time of change unknown, but also the post-change probability model of the observations is often uncertain. When multiple data streams are involved, the number of possible post-change models increases, which slows down detection. This dissertation shows that by employing certain shrinkage estimators, unknown probability models can be estimated in a way that enables faster detection. A practical change detection algorithm should not only detect changes quickly but also allow the user to specify an upper bound on the frequency of false alarms. However, the error metrics traditionally used in the literature do not necessarily capture the quality of decision-making in multi-stream scenarios, where decisions may need to be made simultaneously for different sensors or locations. In this work, decision-making algorithms are developed that allow control of the false discovery rate, the proportion of false alarms among all alarms, a criterion that is both intuitive and scalable. 

The dissertation also examines situations where the change is caused by a spatially propagating phenomenon, such as a propagating radio wave or a natural disaster. The detection of such phenomena is modeled as a change detection problem, and an algorithm is derived that is shown to have certain optimality properties. 

Overall, the theory and methods developed in this dissertation contribute to the design of more efficient, reliable, and interpretable real-time monitoring and decision-making systems.

Keywords: Sequential analysis, quickest change detection, signal processing

Contact:
topi.halme@aalto.fi 

Thesis available for public display 7 days prior to the defence at .