Automatic emotion recognition using portable sensors is gaining attention due to its potential use in real-life scenarios. Existing studies have not explored Galvanic Skin Response and Photoplethysmography sensors exclusively for emotion recognition using nonlinear features with machine learning (ML) classifiers such as Random Forest, Support Vector Machine, Gradient Boosting Machine, K-Nearest Neighbor, and Decision Tree. In this study, we proposed a genuine window sensitivity analysis on a continuous annotation dataset to determine the window duration and percentage of overlap that optimize the classification performance using ML algorithms and nonlinear features, namely, Lyapunov Exponent, Approximate Entropy, and Poincaré indices. We found an optimum window duration of 3 s with $50\%$ overlap and achieved accuracies of 0.75 and 0.74 for both arousal and valence, respectively. In addition, we proposed a Strong Labeling Scheme that kept only the extreme values of the labels, which raised the accuracy score to 0.94 for arousal. Under certain conditions mentioned, traditional ML models offer a good compromise between performance and low computational cost. Our results suggest that well-known ML algorithms can still contribute to the field of emotion recognition, provided that window duration, overlap percentage, and nonlinear features are carefully selected.

Automatic emotion recognition using portable sensors is gaining attention due to its potential use in real-life scenarios. Existing studies have not explored Galvanic Skin Response and Photoplethysmography sensors exclusively for emotion recognition using nonlinear features with machine learning (ML) classifiers such as Random Forest, Support Vector Machine, Gradient Boosting Machine, K-Nearest Neighbor, and Decision Tree. In this study, we proposed a genuine window sensitivity analysis on a continuous annotation dataset to determine the window duration and percentage of overlap that optimize the classification performance using ML algorithms and nonlinear features, namely Lyapunov Exponent, Approximate Entropy, and Poincaré’s indices. We found an optimum window duration of 11 seconds and 16 seconds for valence and arousal, respectively, with 50% overlap, and achieved an accuracy of 0.76 in both dimensions. In addition, we proposed a Strong Labeling Scheme that kept only the extreme values of the labels, which raised the accuracy score to 0.92. Under certain conditions mentioned, traditional ML models offer a good compromise between performance and low computational cost. Our results suggest that well-known ML algorithms can still contribute to the field of emotion recognition, provided that window duration, overlap percentage, and nonlinear features are carefully selected.