Human Activity Recognition Using WISDM: Exploring Class Balancing and ML Techniques

Human Activity Recognition (HAR) is vital in activity sensing research, particularly with wearable sensors. This study addresses dataset imbalance in the WISDM dataset by employing three class balancing techniques: SMOTE, LoRAS, and ProWRA, integrated with five machine learning models, including Decision Tree, Random Forest, K-Nearest Neighbors (K-NN), CatBoost, and XGBoost. Our experimental results demonstrate that LoRAS consistently outperformed the other techniques, achieving a remarkable accuracy of 98.73% with the Random Forest model, significantly higher than SMOTE (97%) and ProWRA (97.12%). Furthermore, LoRAS also excelled in other evaluation metrics, including recall, precision, and F1-score, reinforcing its effectiveness in enhancing model performance for HAR. In contrast, SMOTE and ProWRA exhibited lower performance, indicating that LoRAS is the most effective technique for mitigating class imbalance and improving the accuracy of machine learning models in HAR applications.