Multi-Layer Heterogeneous Ensemble Model for Human Activity Recognition
H S Ganesha, Rinki Gupta, Sindhu Hak Gupta · 2021
Human activity recognition is required in various fields such as sports science, healthcare, elderly care, rehabilitation centers, and other industrial application. In this work, data for five activities of daily living are recorded using three wireless, wearable tri-axis accelerometer sensors. The sensors were placed on the lower limb of the subjects to record multiple observations of the activities for different subjects. First, commonly used machine learning models were tuned to optimize their performance in terms of prediction accuracies. Then, a stacking ensemble was built with the tuned machine learning models as the base learners and three different machine learning models were evaluated when used as metal earners. Further, a novel multi-layer ensemble model was designed by placing the heterogeneous machine learning models in two-stages of meta-learning. The multi-layer stacking improves the prediction performance as compared to the single-layer stacking ensemble as well as the individual classifiers. The average accuracy over all activities obtained using the proposed multi-layer ensemble is 96.1%, which is higher as compared to the other considered models.