A novel approach for identification and classification of human emotions using EEG signals

Abstract

Human affects are intricate combinations of moods that lead to both psychological, physical shifts, evident in facial expressions, body language, and voice modulation. The study of human emotions necessitates the use of emotion models. The study provides an examination of neurophysiological research related to emotions, specifically in the context of valence, dominance and arousal, with EEG signals. The primary objective is to contrast the traditional approaches used in recognizing emotions, including factors such as the quantity and selection of study participants, the abstracted features from data, the choice of classifiers. The review consolidates contemporary methods and their associated results, offering valuable recommendations for researchers seeking to attain reliable and superior outcomes. Emotion classification based on Electroencephalography (EEG) and physiological signals has gained significant attention in recent years due to its potential applications in affective computing and human-computer interaction. In the proposed study, we recommend a novel algorithm that combines a hybrid feature extraction technique with soft labels and weighting factors to improve emotion classification. Our approach incorporates a hybrid technique that combines Fourier Transform and Time Domain features extracted from EEG recordings with existing features of arousal, valence, and dominance from the dataset. To address the overfitting probelm, we employ a manifold learning approach viz. Laplacian Eigenmaps for dimensionality reduction and unsupervised spectral clustering to derive soft labels. These soft labels enhance the generalizability of the classifier. The classification stage employs a Support Vector Machine with a Radial Basis Function kernel, taking into account the soft labels and a weighting factor based on wheel strength. Experimental results demonstrate the effectiveness of our approach, with improved accuracy and specificity compared to a baseline SVM RBF classifier without soft labels...