Emotion Classification Using Optimized Features and Ensemble Learning Techniques for EEG Dataset

Emotion recognition from electroencephalogram (EEG) signals is one of the important real time applications in Brain-Computer Interface (BCI). The proposed research addresses the challenges in emotion classification and analyses different ensemble learning algorithms for classifying emotions, specifically positive, negative, and neutral states, from EEG data. The EEG emotion dataset used in this work is collected from Kaggle, has 2132 samples with 2549 features, where each sample corresponds to EEG signals recorded during various emotional states. To enhance the classification performance, diverse ensemble learning algorithms are applied, including Random Forest, LightGBM, Extra Tree, Gradient Tree, XGBoost, AdaBoost and Bagged Decision Tree. In addition, the optimal set of features are selected through Pearson Correlation Coefficient (PCC) based on threshold value and Mutual Information (MI) based on mi-score. Using the optimal feature set, ensemble learning models are generated and validated using suitable quantitative metrics such as accuracy, sensitivity and specificity for test set. From the results it is observed that Gradient Boosting Tree, XGBoost, Extra Tree achieved above 99% and LightGBM exhibited superior performance of 99.81% accuracy. This outcome proves the effectiveness in using optimal feature set obtained through both PCC and MI techniques.