Interpersonal Neural Synchrony and Subjective Vitality During Social Interaction : An EEG Hyper Scanning Study

Abstract

EEG emotion decoding with high accuracy is an important component of affect-aware HCI, mental health monitoring, and adaptive multimedia. However, many recent methods are based on deep learning models trained on raw signals, which require a lot of data, intensive computation, and sometimes lack interpretability. In this paper, we revisit a traditional but carefully crafted machine learning pipeline and demonstrate that with appropriate feature design and selection, it can achieve state-of-the-art results on multi-class EEG emotion classification. We apply our approach to a real-world EEG dataset consisting of 2,132 trials with 2,548 hand-designed features per trial, including statistical, covariance, eigenvalue-based, entropy, correlation, and FFT features derived from two experimental conditions. After z-score normalization, a Random Forest classifier is first trained on the entire feature set to obtain importance weights, and the top 150 features are selected as a compact affective feature space. A second Random Forest classifier is then trained on this reduced feature space and tested using a stratified hold-out test noteset and 5-fold cross-validation. The proposed approach reaches a test accuracy of 96.72%, with high and well-balanced precision, recall, and F1-score for all three classes of emotions (negative, neutral, positive). The accuracy of the 5-fold cross-validation is 95.87% ± 0.28%, thus confirming the robustness of the approach. The feature importance profiles and the PCA plots also provide insights into which statistical and spectral features convey the most important information about emotions. In summary, the approach shows that an interpretable and feature-driven Random Forest approach can be at least competitive with, if not superior to, more complex deep learning architectures for EEG-based emotion recognition, and that it is computationally efficient and easy to deploy.