CSTO-Net: Cross-View Spectral-Transport and Topology-Operator Network for EEG Emotion Recognition

Multi-view electroencephalography (EEG) emotion recognition must decide how the spectral graph sets of different connectivity views correspond and how their graph propagation operators interact. We present CSTO-Net, a cross-view spectral-transport and topology-operator network for EEG-based emotion recognition. Approach: CSTO-Net integrates three window-wise views: dynamic differential entropy (DDE) as node attributes, amplitude-envelope Pearson correlation coefficient (PCC) graphs, and phase-locking value (PLV) graphs. Cross-View Spectral Transport Routing (CSTR) formulates PCC-PLV spectral correspondence as an entropy-regularized optimal-transport problem with explicit marginal constraints, deliberately separating correspondence estimation from band-importance estimation. Directional Cross-Topology Modulation (DCTM) performs ordered composition of the two routed propagation operators on shared node features and encodes order-dependent propagation discrepancy via a provably skew-symmetric graph-operator commutator with signed decomposition. Main results: On the Database for Emotion Analysis using Physiological Signals (DEAP) dataset, CSTO-Net achieves 93.95 ± 2.50% valence and 94.97 ± 2.72% arousal accuracy under the subject-dependent setting, and 93.73% and 94.54% under participant-wise five-fold subject-independent cross-validation across 32 participants. On the SJTU Emotion EEG Dataset (SEED), it achieves 93.12% under the standard subject-dependent split, 91.36% under 15-fold leave-one-subject-out evaluation across 15 participants, and 98.05 ± 2.32% under subject-dependent five-fold cross-validation. Ablation studies confirm that both CSTR and DCTM contribute substantially to recognition accuracy. Significance: These results demonstrate that explicit cross-view spectral correspondence and topology-operator-level interaction provide complementary benefits for EEG emotion recognition, offering an interpretable route for modeling order-sensitive interactions between functional connectivity views.