The Performance of Feature Selection with the Stability Selection Method in Human Activity Recognition (HAR) Classification Using the Lasso Regression Algorithm

Human activity recognition using wearable sensors has gained significant attention due to its applications in healthcare, smart environments, and personal assistance systems. However, challenges such as high-dimensional data, feature redundancy, and class imbalance remain prominent, affecting the accuracy and interpretability of classification models. This study proposes an ensemble feature selection approach based on Lasso regression to address these issues. The method focuses on selecting the most relevant features from high-dimensional accelerometer and gyroscope data while reducing model complexity. Additionally, stability selection is incorporated to improve the stability and robustness of the feature selection process. The results demonstrate that Lasso regression without stability selection achieves an accuracy of 43.75 %, with low precision and recall values, indicating difficulties in distinguishing between certain activity classes. In contrast, the model with stability selection improves precision (50.60%) and recall (42.75 %), although the overall accuracy decreases slightly to 42.75 %. These findings highlight the effectiveness of stability selection in enhancing feature relevance and addressing class imbalance, while also suggesting the need for further improvements to optimize model performance. The study contributes valuable insights into the application of robust feature selection techniques for human activity recognition tasks in real-world scenarios.