EEG-based Emotion Recognition Using a Hybrid Fusion Approach

In the design of affective computing systems, the use of multiple modalities (physiological and non-physiological) along with fusion strategies is typically considered to enhance emotion recognition. However, such multimodal systems inevitably face increased complexity in both data acquisition and information processing stages. In this work, we propose an emotion recognition system that relies solely on electroencephalogram (EEG) signal features, employing hybrid fusion at both the feature and classification levels. At the first stage, feature fusion is applied to dynamic features extracted from EEG sub-bands across all channels. The resulting feature set is then partitioned into subsets using k-means clustering, with each subset serving as input to an individual classification unit. Finally, the outputs of these classifiers are combined via majority voting fusion, determining the final emotional state. When evaluated on the DEAP dataset using a Support Vector Machine (SVM) classifier, the proposed method achieved 83.8% accuracy in classifying four emotional states. Compared to conventional affective systems using the same dataset with classical classification approaches, our results demonstrate significant improvement in recognition performance.