Gradient Boosting, AdaBoost, and LightGBM for Emotion Recognition: A Comparative Analysis on Wearable Sensor Data

This study investigates how well advanced machine learning algorithms, specifically Gradient Boosting (GB), AdaBoost, and LightGBM, enhance emotion recognition using wearable sensor data. Heart rate and movement data from fifty participants were collected during three scenarios: watching a movie clip before walking, listening to music before walking, and listening to music while walking. Model performance was assessed for classifying binary (happy vs. sad) and multi-class (happy vs. neutral vs. sad) emotions, using metrics such as AUC, F1 score, accuracy, and user lift score. Results indicate that both GB and LightGBM consistently outperform established models like Random Forest and Logistic Regression across all scenarios and both types of emotion classification. Notably, LightGBM achieved the highest accuracies: 91.8% for binary classification and 84.2% for multi-class classification in the 'listening to music while walking' scenario. While AdaBoost's performance was slightly lower than that of Random Forest, it still outperformed the baseline and Logistic Regression models. These findings underscore the potential of GB, AdaBoost, and LightGBM to significantly enhance the accuracy of wearable-based emotion recognition systems.