EEG-Based Emotion Classification Using Deep Capsule Networks for Subject-Independent and Dependent Scenarios

Emotion recognition from electroencephalography (EEG) signals is an important component of emotionally intelligent human–computer interaction systems. However, existing approaches often rely on handcrafted features or conventional deep learning architectures that struggle to generalize across subjects due to the high variability of EEG signals. Most prior studies focus primarily on binary emotion classification and provide limited investigation of more complex multi-class scenarios. This work presents EmoCaps, an end-to end deep learning framework based on a capsule network with a self-attention–guided routing mechanism to learn discriminative representations directly from raw EEG signals. The proposed model is evaluated on the DEAP dataset under both subject dependent and subject-independent settings and across binary and multi-class emotion recognition tasks involving valence, arousal, and dominance dimensions. Experimental results demonstrate that EmoCaps consistently outperforms several representative deep learning models. In subject-dependent binary classification, the proposed approach achieves accuracies of up to 97%, while in the more challenging subject-independent setting it exceeds 80% accuracy across emotional dimensions. The framework also achieves strong performance in four-class and eight-class emotion recognition tasks and provides, to the best of our knowledge, the first reported results for subject independent multi-class emotion recognition on this dataset. Although the capsule-based architecture introduces higher computational cost, the proposed model significantly improves robustness and generalization across subjects. These results highlight the potential of EmoCaps for real-world emotion-aware applications in healthcare, adaptive learning, and affective computing systems.