Emotion Recognition Classification with Differential Entropy and Power Spectral Density Features

Feature selection is crucial for emotion recognition using machine learning algorithms. This has been extensively applied in various domains. This study utilized Electroencephalography (EEG) signals of positive, neutral-polar, and negative emotions produced by 15 individuals while watching 15 movie clips within the SEED database. Specifically, Differential Entropy (DE) and Power Spectral Density (PSD) were extracted as features, and signal smoothing was conducted on the extracted features using both Linear Dynamic Systems (LDS) and traditional moving average techniques. The performance of four classifier models, SVM-RFE, logistic regression (LR), random forest (RF), and multilayer perceptron (MLP), was evaluated using accuracy (ACC), precision (PRE), recall (REC), and F1 scores. Among them, SVM-RFE and LR had the best accuracy, in SVM-RFE and LR, DE had 87.4% and 85.9% accuracy under LDS method; PSD had 63.70% and 70.37% accuracy under LDS method. The study’s findings indicate that DE feature classification outperforms PSD feature classification significantly, and the accuracies obtained using LDS are generally higher than those obtained using the traditional moving average method.