Hybrid feature selection technique for classification of human activity recognition
Through the advancement of wearable sensors, wireless communication, and machine learning techniques, Assistive Technologies (AT) which endorse autonomous, active, and healthy lifestyles are emerging in recent years. Among these advances, Human Activity Recognition (HAR) is one of the most innovativ...
No Tags, Be the first to tag this record!
|Through the advancement of wearable sensors, wireless communication, and machine learning techniques, Assistive Technologies (AT) which endorse autonomous, active, and healthy lifestyles are emerging in recent years. Among these advances, Human Activity Recognition (HAR) is one of the most innovative means to support or monitor older people in their daily lives. Extensive research in the field of HAR is also essential for the enhancement of the quality of a person"s health. However, misclassifications such as intra-class variation and inter-class overlap in similar activities degrade classification accuracy. To improve the recognition of daily human activities, handcrafted features of time-domain and frequency-domain are combined. However, several extracted features may not be significant in describing the activities. Therefore, this research aims to propose a hybrid feature selection technique for optimal human activities recognition. The methodology proposed for this research is the Ensemble Filter (Relief-F and mRMR) to select the most relevant and less redundant features. Although a filter feature ranking approach is commonly used in related studies, most works fail to consider the threshold limit to exclude unnecessary and redundant features. The hybrid Binary Harmony Search-Artificial Bee Colony (BHS-ABC) algorithm further evaluated the quality of the selected features in this research. Two HAR datasets using accelerometer and gyroscope sensors from the smartphone device were evaluated, covering various daily human activities. An ensemble Random Forest (RF) was used as the base classifier to evaluate the performance of the hybrid algorithm. The performance was then compared with other shallow machine learning techniques such as Support Vector Machine (SVM), k-Nearest Neighbour (kNN), and Naïve Bayes (NB). One-versus-All (OvA) class binarization technique was adopted to enhance the detection of high-class correlations by converting initial problems of multiple classifications into a set of two-class classification problems. The proposed hybrid of Ensemble Filter with BHS-ABC for this research is capable of obtaining optimum feature subsets with high accuracy. The average of SBHAR and USCHAD datasets accuracy have improved by 3.2% and 1.6%, respectively, from the benchmark results.