M. Faris Al Hakim, Budi Prasetiyo, Regina Ayumi Ulayyaa
This study explores the impact of various sampling techniques on the performance of machine learning models for human activity recognition. The sampling methods evaluated include SMOTE, Random Over Sampling (ROS), SMOTE-Tomek, and CBSO, and their effects on models such as Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), and XGBoost (XGB) are examined. The results demonstrate that sampling methods significantly improve model performance, with RF showing the improvement, particularly with CBSO (accuracy: 93.08%). Logistic Regression using SMOTE-Tomek achieved the highest accuracy of all methods used with acucracy of 96.34%, followed by SVM with CBSO (96.06%), and XGBoost with ROS (94.13%). These performances were achieved even without model hyperparameter tuning or feature selection. The findings highlight the importance of selecting appropriate sampling methods for addressing class imbalance and optimizing model performance in human activity recognition tasks. © 2025 IEEE.
Department of Computer Science, Universitas Negeri Semarang, Semarang, Indonesia
