Wrapper Filter Approach for Accelerometer-Based Human Activity Recognition

With the widespread use of mobile devices all over the world, a new interesting and challenging research area known as Activity Recognition (AR) with many application domains is evolved. Basically, activity recognition aims to identify certain physical activities such as walking, jogging, sitting, standing, etc., performed daily by humans. In this paper, we investigated the effectiveness of wrapper-based feature selection approach for accelerometer-based human activity recognition. Our approach utilizes Sequential Forward Selection (SFS) technique based on three machine learning algorithms: Random Forest (RF), K-Nearest Neighbor (K-NN), and Gradient-Boosted Tree (GBT). A standard and publicly available dataset called WISDM (Wireless Sensor Data Mining), which contains accelerometer-based time series data collected from thirty-six volunteers, was used for performance evaluation of the proposed model. The experimental results showed that our GBT-based recognition model outperforms previously suggested solutions and establishing state-of-the-art performance for this dataset.