Evaluating Classical Machine Learning Algorithms for Multimodal Human Activity Recognition

Multimodal human activity recognition (HAR) is becoming increasingly critical in applications such as healthcare, sports analytics, and smart environments. While deep learning algorithms have dominated recent research, classical machine learning methods remain highly effective for classifying activities using multimodal data, particularly in scenarios with limited computational resources. This study leverages LazyPredict for rapid model benchmarking and Optuna for hyperparameter tuning to accelerate the development of efficient classifiers. Ensemble learning models specifically were identified as the most effective for both intrapersonal and interpersonal activity recognition tasks. Results indicate that Light Gradient Boosting Machine (LGBM) and Random Forest classifiers perform best, achieving accuracies of approximately 97% and 82% for intrapersonal and interpersonal models, respectively. Moreover, the hyperparameter tuning using Optuna improves accuracy but significantly increases model sizes, indicating that additional size constraints are needed for edge deployment. Important features were also highlighted in the study showing that electrodermal activity (EDA) from the wrist and y-axis accelerometer data from the wrist is the most important feature for intrapersonal and interpersonal model, respectively. Overall, this study demonstrates practical methods for tuning classical machine learning algorithms for resource-constrained devices, making them suitable for real-world deployment in edge computing and IoT applications.