A Comparative Analysis of Feature Selection Approaches for Sensor-Based Human Activity Recognition

In the era of infrastructure-less sensing, human activity recognition takes a new leap by exploiting the ubiquity of smartphone sensors. However, the limited computational capability of the smart handhelds hampers the inherent need for real-time responsiveness of the application. Thus, dimensionality reduction through Feature Selection (FS) techniques could be a precursor for on-device learning and prediction of activities. Most of the existing FS approaches focus on the algorithmic perspectives and ignore the data-intensive nature of the application. So, in this paper, we perform a comparative study of the FS techniques utilized for HAR subject to the role of feature preprocessing. Apart from detailing the FS methods, the work also shows how to apply those methods in combination. The experimental results across the benchmark datasets show the efficiency of our application strategies of the FS techniques.