*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Change Detection in Time Series Analysis and Machine Learning
Committee:
Dr. Davenport, Advisor
Dr. Romberg, Chair
Dr. Xie
Abstract: The objective of the proposed research is to use change detection for machine learning tasks in streaming data settings. This objective focuses on two major themes. Firstly, I propose improvements to existing change detection procedures so that multiple change points in streaming data can be detected in an improved manner. The second theme explores how change detection can provide information about data and class distributions which can be exploited by different machine learning models. For improved multiple sequential change detection, I propose a symmetrical change detection procedure which has a similar power to identify changes between distributions. This means that if given two distributions A and B, the change detection power would be similar for a change from distribution A to distribution B and for a change from distribution B to distribution A. I also propose semi-supervised change detection procedures which could utilize limited class labels for improved change detection. For the second theme, I propose that change detection itself can help identify instances where class distributions change. This can help provide auxiliary similar/dissimilar information in an unsupervised manner which can be used in conjunction with available labels to devise a semi-supervised sequence classification model. Real world settings also involve data and class distribution shifts which can cause supervised models to perform poorly. I explore how change detection can provide useful information on structural changes within streaming data so that supervised models can be retrained while requiring fewer training samples
