Dynamic ensemble learning model for Human Activity Recognition using wearable sensors

Human Activity Recognition (HAR) is a highly significant topic within the context of IoT applications. Recent state-of-the-art methods of HAR have been designed based on classic ensemble learning approaches and deep learning techniques. However, those models suffer from several limitations, such as class imbalance issues and high computational cost. We designed a dynamic ensemble classification model integrated with Dynamic Mode Decomposition (DMD) to recognise human activity. The proposed model is designed to extract the intrinsic temporal dynamics of sensor signals while simultaneously leveraging the diversity of multiple base classifiers. Unlike classic static ensembles, the proposed ensemble model uses a three-phase architecture involving of base classifier generation, competence-based dynamic selection, and adaptive aggregation to improve recognition performance. By incorporating both single-label accuracy and ranking loss, the proposed ensemble model ensures that only the most relevant classifiers contribute to the final decision, thereby improving accuracy and reducing the influence of weak learners. The performance of the proposed model is verified using three datasets: UCI-HAR, Opportunity and WISDM. The developed model achieved an impressive accuracy of 96, 97, 98% and F1-score of 98, 96, 97%for UCI-HAR, Opportunity, and WISDM, respectively. Our results show that the proposed model performs better than its counterparts in the literature.