A GRU+KAN Hybrid Deep Learning Framework for EEG-Based Emotion Recognition

Emotion recognition from EEG signals is a fundamental task in affective computing, with growing applications in healthcare, adaptive systems, and human-computer interaction. Recent deep learning models have utilised Convolutional Neural Networks (CNNs) for spatial feature extraction and Recurrent Neural Networks (RNNs) for temporal modeling. However, these models often fail to leverage informative representations embedded within intermediate layers, leading to suboptimal classification performance. To address these limitations, this study proposes a novel hybrid architecture-GRU+KAN-that combines the temporal modeling strength of Gated Recurrent Units (GRUs) with the expressive nonlinear approximation capabilities of Kolmogorov-Arnold Networks (KANs). Unlike standard RNNs, GRUs are designed to model long-term dependencies with reduced computational overhead. Meanwhile, KANs offer a powerful mechanism to learn complex, multivariate relationships through adaptive spline-based activations. Wavelet-based denoising using the Daubechies db4 filter is applied as a preprocessing step to improve EEG signal quality. Experiments conducted on the GAMEEMO and LUMED datasets demonstrate that GRU+KAN significantly outperforms the baseline GRU model with gains exceeding 15% on GAMEEMO and 6% on LUMED. The proposed model achieved 91.13% accuracy and 91.82% F1-score on GAMEEMO, and 89.97% accuracy and 90.19% F1-score on LUMED, all while reducing parameter count by nearly 50%.