Boosting EEG-Based Emotion Recognition with Kolmogorov-Arnold Networks

As biomedical signal processing continues to advance, the need for reliable and deployable solutions for identifying human emotional states grows. Electroencephalography (EEG) is a promising modality for emotion recognition; however, existing systems often face challenges related to accuracy and efficiency. In this work, the use of the Kolmogorov-Arnold Network (KAN) for EEG-based emotion recognition is studied. KANs are a novel type of neural network that utilize edge-based activation functions, learning complex patterns with fewer parameters than traditional Multi-Layer Perceptrons (MLP). In addition, to enhance the quality of EEG signals, wavelet denoising is utilized. The proposed KAN model is evaluated on two publicly available datasets: GAMEEMO and LUMED. The results demonstrate that KAN significantly outperforms MLP in both accuracy and F1 score. On the GAMEEMO dataset, KAN achieved an average accuracy of 80.79% and an F1 score of 78.63%, while on the LUMED dataset, it reached an accuracy of 90.95% and an F1 score of 89.53%. This is a significant improvement of over 4% in both metrics compared to MLP. These results highlight the potential of KANs as an efficient and effective approach for EEG-based emotion recognition.