Electroencephalogram based human emotion classification for valence and arousal using machine learning approach

Abstract

Humans have unique ability to express emotions and electroencephalogram (EEG) signals are one of the sought-after ways to analyze a person’s emotional state. However, extracting proper emotion related features from EEG and finding corresponding emotion is challenging because of complex nature of emotions and underlying brain activities. The objective of this paper is to address this issue for more accurate emotion classification based on EEG. It also compares feature extraction methods namely fast fourier transform (FFT) and discrete wavelet transform (DWT). DEAP dataset is used for classification of human emotions through support vector machine (SVM) and K-nearest neighbor (KNN) algorithms by considering features such as standard deviation, mean, variance, power spectrum density (PSD) for FFT; and energy, entropy for DWT. It is observed that feature extraction from FFT yielded better results than DWT and KNN gave more accuracy of 96.61% for valence and 96.42% for arousal as compared to SVM. The proposed method based on PSD and FFT fared better than other existing ones in terms of accuracy when compared against different features and feature extraction techniques. This approach is expected to help researchers to understand feature extraction from EEG signals and decide proper features and techniques for better implementation.