Redundancy-Aware Feature Selection using mRMR and F-Test for EEG Emotion Classification

Abstract

Emotions play an essential role in human interaction, driving the development of reliable automatic emotion recognition systems. Electroencephalography (EEG) offers a noninvasive method to record neural activity related to emotional states; however, many existing studies focus on limited feature configurations or binary classification problems. This research examines the influence of feature dimensionality and classifier selection on three-class EEG-based emotion recognition involving positive, neutral, and negative categories. The primary contribution of this study is a systematic assessment of feature and classifier compatibility across 28 experimental scenarios within a unified evaluation framework. Using a publicly available EEG dataset containing statistical and spectral features, selection was conducted using F-test and Minimum Redundancy Maximum Relevance (mRMR) methods, isolating the top 5, 10, and 15 features alongside the complete set. Four classifiers (Random Forest, Support Vector Machine, K-Nearest Neighbors, and Neural Networks) were evaluated via a 70/30 hold-out validation scheme using accuracy, F1-score, and Area Under the Curve (AUC). Results indicate that Random Forest trained with the full feature set achieved the highest performance, reaching 99.53% accuracy and 0.9994 AUC. These findings suggest that ensemble-based models demonstrate greater robustness when handling high-dimensional EEG features in multi-class emotion recognition.