A Comparative Study on Missing Data Handling Using Machine Learning for Human Activity Recognition

In mobile computing era, human activity recognition is an important research area for eldercare and healthcare center. Activity recognition performance decrease while partial data are lost due to various reasons, such as power limitation, sensors data transmission capacity, hardware failures, and network issues (e.g., packet collision, unreliable link, and unexpected damage). Therefore, data loss or missing is a realistic challenge. In this research, we evaluate the performance of algorithms to perform their ability to handling missing data. It is required to investigate the optimal missing percentage in the dataset and its effect on accuracy. In this case, missing data are implemented randomly in the dataset with incremental scaling. We evaluate the performance using two benchmark datasets named HASC dataset and Single Chest-Mounted accelerometer dataset with data missing at random pattern. Here, we classified using support vector machine and random forest classifiers with eleven statistical features. Initially in the original dataset, there are no missing data in the dataset. However, to evaluate the performance, we added various levels of missing data randomly and studied the performances. Our study demonstrated that over 8% missing data in the dataset greatly reduce recognition results. Moreover, this recognition result has impact on sampling rate during data collection time.