Personalization in Mobile Activity Recognition System Using K-Medoids Clustering Algorithm

Abstract

Nowadays mobile activity recognition (AR) has been creating great potentials in many applications including mobile healthcare and context-aware systems. Human activities could be detected based on sensory data that are available on today's smart phone. In this study, we consider mobile phones as an independent device since sending the data to central server can generate privacy issues. Furthermore, applying AR on mobile phone does not only require an effective accuracy rate but also the lowest power consumption. Normally, an AR model learnt from acceleration data of a specific person is distributed to other people to recognize the same activities instead of generating different models individually. This work often cannot create accurate results on the prediction in broad range of participants. Moreover, such AR model also has to allow each user to update his new activities independently. Therefore, we propose an algorithm that integrates Support Vector Machine classifier and K-medoids clustering method to resolve completely the demand.