Networked and Embedded Systems Laboratory

Tryear

Trmonth

Trnumber

Title

Abstract

Mobile phones’ increasing ubiquity has created many opportunities for personal context sensing. Personal activity is an important part of a user’s context, and automatically recognizing it is vital for health and fitness monitoring applications. Recording a stream of activity data enables monitoring patients with chronic conditions affecting ambulation and motion, as well as those undergoing rehabilitation treatments. Modern mobile phones are powerful enough to perform activity classification in real time, but they typically use a static classifier that is trained in advance or require the user to manually add training data after the application is on his/her device. This paper investigates ways of automatically augmenting activity classifiers after they are deployed in an application. It compares active learning and three different semi-supervised learning methods, self-learning, En-Co-Training, and democratic co-learning, to determine which show promise for this purpose. The results show that active learning, En-Co-Training, and democratic co-learning perform well when the initial classifier’s accuracy is low (75-80%). When the initial accuracy is already high (90%), these methods are no longer effective, but they do not hurt the accuracy either. Overall, active learning gave the highest improvement, but democratic co- learning was almost as good and does not require user interaction. Thus, democratic co-learning would be the best choice for most applications, since it would significantly increase the accuracy for initial classifiers that performed poorly.

Filename

File

Urlpdfpaper

Urlsrcpaper

Urlpdfpresentation

Urlsrcpresentation

Urlavmedia

Urldoi

Urlpublisher

Urlgooglescholar

Urlciteseer

Pubin

Pubvol

Pubnum

Pubnum end

Pubpagefirst

Pubpagelast

Pubpagecount

Pubdate

Pubdate end

Pubplace

Publisher

Ispublic

Islabdocument

Miscattributes

Document category

Main research area

Editing document