Autoencoder Based Nonlinear Feature Extraction from EDA Signals for Emotion Recognition

The field of emotion recognition plays a pivotal role in enhancing the quality and effectiveness of human-computer interaction. Electrodermal Activity (EDA) signals, which reflect autonomic nervous system response, have been widely employed to understand user emotional states. However, the inherent nonlinear and non-stationary nature of these signals poses a challenge for traditional analysis methods. To address these challenges, recent studies have focused on non-linear signal processing methods. This work proposes a novel approach utilizing deep learning based Autoencoders for the extraction of non-linear features from EDA signals. For this, first, we estimate the phasic component of EDA signals that are obtained from a public dataset. Then we compute the 64 bottleneck features using Autoencoders to effectively capture the non-linear characteristics of EDA signals. Further, the RF-RFE method is employed for the selection of the top five most informative features. These selected features are then fed into machine learning algorithms including random Forest, support vector machine, Adaboost, and extreme gradient boosting for the classification of arousal and valence dimensions. Our results demonstrate the robustness of this approach and its ability to accurately discriminate complex emotional states. This work significantly advances EDA-based emotion recognition and opens avenues for more emotionally attuned human-computer interaction systems.