Parameter Optimization for Online Change Detection in Activity Monitoring Using Multivariate Exponentially Weighted Moving Average (MEWMA)

In recent years, body worn sensors have become popular for the purpose of activity recognition. The sensors used to capture a large amount of data in a short period of time which contain meaningful events. The change points in this data can be used to specify transition to a distinct event which can subsequently be used in various scenarios such as to identify changes in patient vital signs in a medical domain or to assist in the process of generating activity labels for the purposes of annotating real-world datasets. A change point can also be used to identify the transition from one activity to another. The multivariate exponentially weighted moving average (MEWMA) algorithm has been proposed to automatically detect such change points for transitions in user activity. The MEWMA approach does not require any assumptions to be made in relation to the underlying distributions to evaluate multivariate data streams and can run in an online scenario. The focus of this paper is to evaluate the performance of the MEWMA approach for change point detection in user activity and to determine the optimal parameter values by tuning and analyzing different parameters of MEWMA. Optimal parameter selection results in an algorithm to detect accurate change points and minimize false alarms. Results are presented and compared based on real world accelerometer data for standard and optimal parameters evaluated using different metrics such as accuracy, precision, G-mean and F-measures.