Work

TFDCA for EEG emotion recognition through dynamic channel aggregation

Year: 2026
Type: preprint
Abstract: Electroencephalography (EEG) provides valuable information for emotion recognition owing to its high temporal resolution and sensitivity to emotional states. However, existing brain region methods often rely on predefined static partitions, which have limited capacity to capture dynamic channel interactions, functional overlap, and reorganization across regions. A temporal–frequency fusion method with dynamic channel aggregation (TFDCA) is proposed to characterize channel dependencies across emotional states. A fusion encoder extracts complementary temporal and frequency features: lightweight convolutional attention weights multiband differential entropy features, while gated recurrent units capture temporal dependencies. Adaptive weighting integrates these features into multiscale node representations. Channel similarities derived from self-attention are combined with electrode spatial topology under a sparsity constraint to construct an adjacency matrix. Differentiable soft sorting jointly reorders its rows and columns to aggregate similar channels into regions, incorporating local structure and global channel relationships. A brain region transformer further captures interregional dependencies for emotion classification. On SEED, TFDCA achieved accuracies of 96.96% and 85.31% under subject-dependent and subject-independent protocols, respectively. On DEAP, subject-dependent accuracies were 92.84% for valence and 93.02% for arousal, while subject-independent accuracies were 65.13% and 66.24%, respectively. Visualization analyses revealed emotion-related channel aggregation patterns and their variation, providing evidence for interpreting the learned representations. These findings support dynamic channel aggregation as an approach to representing EEG channel relationships for emotion recognition.
Language: English
Open Access status: green
OpenAlex ID: W7225867333
Cites: 39
Cited by:
Related to: 10