Ensem-EnsemHAR: Human Activity Recognition using Ensemble of Ensemble Methods and Smartphone Sensor Data

• Published in: International Conference on Computing, Communication, and Networking Technologies (Online) pp. 1 - 7
• Main Authors: Islam, S.M. Mohidul, Talukder, Kamrul Hasan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.06.2024
• Subjects: Accelerometers; Accuracy; Data models; Ensemble learning; Ensemble of ensembles; Grid Search; Human activity; Human activity recognition; Medical services; Random Search; Robustness; Security; Smartphone sensor; Training; Tuning
• ISSN: 2473-7674
• DOI: 10.1109/ICCCNT61001.2024.10723972

Recognizing human activity has many applications in real life such as in biometrics, military, security, medical care applications, etc. In this paper, we proposed a new activity recognition method using the voting fusion of various ensemble methods. The data are subject-wise split into training and test sets to accomplish a person-independent recognition model. As base ensemble methods we have used random forest, extra-trees, AdaBoost, Histogram-based Gradient Boost, and XGBoost classifier. The hyper-parameters of base ensemble methods are fine-tuned using four different methods: RandomizedSearchCV, HalvingRandomSearchCV, GridSearchCV, and HalvingGridSearchCV. The recognition accuracies of each base ensemble for various parameter tuning techniques as well as for scikit-learn default values of the parameters are compared and we selected those values of hyper-parameters for which we obtained the highest recognition accuracy. Then the base ensembles are used in a voting ensemble which provides the final recognition result. The prediction results show the robustness of the proposed method to many similar existing methods, especially for unbalanced datasets.