Quaternion Spiking Attention Network for Cross-Subject EEG Emotion Recognition

Spiking Neural Networks (SNNs) provide sparse event-driven computation and are therefore attractive for resource-constrained EEG applications. We introduce Q-SNet, a quaternion-based framework for cross-subject EEG emotion recognition, to enhance the modeling capability of SNNs through structured hypercomplex representations. Specifically, Q-SNet employs the quaternion-valued coupling to encode multi-channel EEG signals into quaternion representations. A quaternion rotation attention module is introduced, where learnable quaternion rotations are applied to the value components. The rotation-aligned features are then processed by a Quaternion Leaky Integrate-and-Fire (Q-LIF) neuron, in which spike firing is triggered by the quaternion magnitude, leveraging its rotation-invariant property to promote more consistent firing behavior under subject-dependent geometric perturbations. Q-SNet achieves improved cross-subject robustness with a compact parameterization while retaining the event-driven characteristics of SNNs. Extensive experiments on SEED, SEED-IV, SEED-V and HBUED demonstrate the effectiveness of Q-SNet, achieving accuracies of 93.76%, 78.25%, 89.98% and 59.69%, respectively. The code can be found at https://github.com/liuyici/Q-SNet.