Linear–Nonlinear Feature Reconstruction Network for Emotion Recognition From Brain Functional Connectivity

The electroencephalogram records both the static performance and dynamic changes in the brain electrical activity, with linear and nonlinear features capable of modeling these static and dynamic aspects. However, traditional EEG-based emotion recognition methods rely excessively on linear features (static analysis) without fully considering nonlinear features (dynamic changes). This limitation hinders a comprehensive capture of the complex interactions and dynamic variations between brain regions. This study proposes an EEG emotion recognition method based on a linear-nonlinear feature reconstruction network using the brain functional connectivity. The method comprehensively captures the complex interactions and dynamic changes between brain regions by extracting linear and nonlinear features from the dynamic and static connectivities. To enhance the feature discriminability, a feature self-reconstruction encoder (FSRE) is introduced, which refines the feature representation and improves the accuracy and discrimination of features. A feature interaction and reconstruction (FIR) module uses a bidirectional reconstruction strategy to interact with and reconstruct features based on their relationships, thereby enhancing the correlation between features and target classification. Experiments conducted on the SEED and SEED-IV datasets validate the effectiveness of this method, demonstrating that it achieves higher accuracy and robustness in emotion classification tasks compared with other advanced methods.