Multiscale Efficient Attention-Based EEG Emotion Classification with Random Feature Decoding and Reconstruction Strategy

This study introduces a multi-scale efficient attention-based EEG emotion recognition approach, The proposed method employs RFDR for stochastic feature decoding and feature-dimension-based reconstruction, ensuring that the reconstructed features are solely based on feature correlations. Simultaneously, a deep separable convolutional network is utilized to construct a multi-scale feature extraction backbone, designed to identify emotionspecific feature information while reducing computational complexity. An Efficient Multi-Scale Attention (EMA) module is incorporated to enhance the weighting of feature values within each channel and facilitate cross-spatial information aggregation. The efficacy of this method was evaluated on the HIED database, achieving an average accuracy of 86.87 % in a six-class emotion recognition task, and on the SEED database, attaining an average accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 6. 6 4 \%}$</tex> in a three-class emotion recognition task.