Weighted Heterogeneous Recurrence Network Analysis achieved an accuracy of 0.8418 (AUC 0.9093) on the SEED database and 0.7979 (AUC 0.8011) on SEED-IV for EEG-based emotion recognition.
WHRNA provides an interpretable nonlinear-dynamical framework for EEG-based emotion recognition with competitive performance compared to representative methods.
Electroencephalography (EEG)-based emotion recognition has potential for healthcare informatics and affective computing, but remains challenging because EEG signals are nonlinear, nonstationary, and highly dynamic. This study presents Weighted Heterogeneous Recurrence Network Analysis (WHRNA), an interpretable recurrence-network framework that characterizes edge-level recurrence strength and recurrent relationships within and across heterogeneous subspaces of the embedded EEG state space. EEG signals from the SEED and SEED-IV databases were decomposed into beta, gamma, and high-gamma bands, embedded into a low-dimensional state space using Uniform Manifold Approximation and Projection (UMAP), and segmented into Voronoi-defined subspaces. WHRNA then quantified Global-WHRN, Inner-WHRN, and Cross-WHRN recurrence structures to form graph-based feature representations. After multi-model importance-based feature selection, supervised classifiers were evaluated under repeated One-vs-All classification. LightGBM achieved the best performance on SEED (accuracy = 0.8418, AUC = 0.9093), while XGBoost achieved the best performance on SEED-IV (accuracy = 0.7979, AUC = 0.8011). Although some recent high-capacity deep learning models report higher accuracy, WHRNA provides competitive performance against several representative methods while offering an interpretable nonlinear-dynamical view of EEG emotion recognition. Sensitivity analyses show that SEED is mainly characterized by high-gamma betweenness-related recurrence structures involving a dominant subspace, whereas SEED-IV requires more distributed multi-band recurrence evidence across subspaces. These findings suggest that WHRNA can support interpretable EEG-based emotion analysis for affective computing and healthcare-oriented applications.
Wang et al. (Wed,) conducted a other in Emotion recognition. Weighted Heterogeneous Recurrence Network Analysis (WHRNA) vs. Representative methods and deep learning models was evaluated on Classification accuracy and AUC. Weighted Heterogeneous Recurrence Network Analysis achieved an accuracy of 0.8418 (AUC 0.9093) on the SEED database and 0.7979 (AUC 0.8011) on SEED-IV for EEG-based emotion recognition.