Optimizing Machine Learning Models for Human Activity Recognition via Feature Engineering and Selection

This paper explores human activity recognition (HAR) using machine learning models to classify activities based on sensor data with high precision. The study leverages the UCI HAR dataset, encompassing six physical activities (WALKING, WALKING UPSTAIRS, WALKING DOWNSTAIRS, SITTING, STANDING, and LAYING). Our study follows a three-stage approach: raw dataset analysis, feature engineering, and feature selection. Twelve machine learning models were initially evaluated on the raw UCI HAR dataset to establish baseline performance. Feature engineering in the second stage enhanced model accuracy for several classifiers. In the final stage, feature selection techniques, including ANOVA, Chi-Squared, Forward Selection, Recursive Feature Elimination (RFE), Random Forestbased selection, and Decision Tree-based selection, were applied to refine the dataset further and improve training efficacy. The accuracy, precision, recall, and F1 score assessed model performance. The results demonstrate that feature engineering and selection significantly enhance HAR classification accuracy and computational efficiency. Linear SVC, Logistic Regression, and Artificial Neural Network (ANN) consistently outperformed other models, underscoring the importance of data preparation and feature selection in HAR applications. These findings provide valuable insights for developing optimized human activity monitoring systems.