Evaluation of semi-supervised clustering and feature selection for human activity recognition

A lot of concern is shifted nowadays toward human activity recognition for developing powerful systems that assist numerous humans such as patients and elder people. Such smart systems automatically recognize human activities by learning the Activities of Daily Living (ADLs) and then making a suitable decision. Activity recognition systems are currently employed in developing many smart technologies (e.g., smart homes) and their uses have been dramatically increased with availability of Internet of Things (IoT) technology. Researchers have used various machine learning techniques for developing activity recognition systems. Nevertheless, there are some techniques have not been sufficiently exploited in this research direction. In this work, we present a framework to evaluate performance of one of these techniques. The presented technique is based on employing semi-supervised clustering and feature selection for grouping data collected by sensors located in smart homes. Employing semi-supervised clustering technique decreases the need for preparing a huge amount of labelled data that is required for learning activity recognition systems. Additionally, the presented technique improves performance of data clustering by decreasing a risk of grouping data into clusters that do not correspond to targeted activities. Moreover, we take into consideration importance of decreasing computational time complexity for making the presented technique applicable to smart systems located in homes. We conducted various experiments to evaluate performance of employing semi-supervised clustering and feature selection for human activity recognition by using partially labelled data. Experiment results have shown that the presented technique provides remarkable accuracy.