Optimizing Human Activity Recognition for Resource-Constrained Environments Using Random Forest and Feature Subset Selection

Optimizing Human Activity Recognition (HAR) for devices with limited resources is essential for broader use. HAR is becoming a vital component in the design of intelligent health monitoring systems, smart living environments, and behavioraware mobile applications. Despite recent progress in machine learning, deploying such systems on devices with limited resources like smartphones and wearables presents serious limitations related to energy use, memory constraints, and processing capability. This paper presents a new optimization method that takes advantage of the interpretability and strong performance of the Random Forest algorithm. By using Random Forest’s feature importance, we reduce the complexity of smartphone sensor data while carefully measuring the balance between the number of features and HAR accuracy. By evaluating the importance of each feature through the algorithm’s internal structure, we selectively retain only the most informative features from smartphone sensor data. This results in a substantial reduction in data dimensionality while maintaining strong classification performance. Our experiments on a benchmark dataset show that we can achieve over $90 \%$ feature reduction with little impact on performance (for example, a $\sim 0.34 \%$ accuracy drop using only the top 50 features). This approach greatly improves computational efficiency, lowers memory usage, and extends battery life. It provides a practical and easy-to-understand framework for running high-performance HAR on devices. The findings also provide valuable insight into which sensor signals are most useful for distinguishing between different types of activities. This contributes to a more interpretable and resource-aware framework for future applications of activity recognition in real-world scenarios.