Optimizing feature selection in multilayer ensemble models for improved HAR accuracy

Human activity recognition (HAR), is an increasingly significant research area as it can be used in many fields of application such as; health care, elderly monitoring, sports training, and smart homes. In this research we developed a novel multi-layer ensemble model based on a combination of a genetic algorithm (GA) to optimize feature selection and hierarchical learning to solve the issues of high dimensional data, feature redundancy and over fitting in HAR. Our model systematically reduces the number of features required to recognize activities while maintaining the most important features; thus, allowing the base learner to learn patterns across multiple layers. We demonstrated through experiments using three standard benchmark datasets-UCI HAR, WISDM, and PAMAP2, that our method significantly outperformed standard methods achieving 96.8% accuracy, and reduced the amount of feature sets by more than 70%. Evaluation metrics including; precision, recall, F1-score, and ROC-AUC, further validated the robustness of our model; while statistical tests confirmed the improvement in performance. Additionally, our framework improved the efficiency and interpretability of our model, which will enable it to be practically implemented in real time environments. These results demonstrate the potential of combining feature selection optimized by a GA and hierarchical ensembles in HAR, and provide avenues for future work in cross domain adaptability and multimodal HAR systems.