An asynchronous multi-view learning approach for activity recognition using wearables

In this paper, we introduce an Asynchronous Multiview Learning (AML) approach to allow accurate transfer of activity classification models across asynchronous sensor views. Our study is motivated by the highly dynamic nature of health monitoring using wearable sensors. Such dynamics include changes in sensing platform (e.g., sensor upgrade) and platform settings (e.g., sampling frequency, on-body sensor location), which result in failure of the machine learning algorithms if they remain untrained in the new setting. Our approach allows machine learning algorithms to automatically reconfigure without any need for labeled training data in the new setting. Our evaluation using real data collected with wearable motion sensors demonstrates that the average classification accuracy using our automatically labeled training data is 85.2%. This accuracy is only 3.4% to 4.5% less than the experimental upper bound, where ground truth labeled training data are used to develop a new activity recognition classifier.