tsta

فارسی

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Graph Attention Networks for EEG-Based Emotion Recognition: Focus on Channel‑Level Attention

Authors :

Akbar Asgharzadeh-Bonab¹ Hamid Bigdeli² Mohammad Javad Heidari³

1- دانشگاه فرماندهی و ستاد آجا 2- دانشگاه فرماندهی و ستاد آجا 3- دانشکده برق‌و‌کامپیوتر، دانشگاه تهران

Keywords :

EEG،emotion recognition،graph attention networks،channel-level attention،differential entropy،functional connectivity،DEAP

Abstract :

We investigate the specific contribution of channel-level attention to EEG-based emotion recognition by introducing Channel-GAT, a compact two-layer graph attention network that operates on the 10–20 sensor graph. Each 1-s window is represented as a 32-node graph with differential-entropy band features; edges capture spatial proximity and, when reliable, functional connectivity. On the DEAP dataset, Channel-GAT consistently outperforms non-attention graph models and common grid-based baselines while remaining lightweight, reaching around 93% accuracy for valence and 92% for arousal in subject-dependent evaluation and competitive performance under leave-one-subject-out testing. Ablations show that attention yields a clear gain over a plain GCN, and that using an explicit graph structure substantially improves over MLPs without topology. Attention maps emphasize frontal regions for valence and parietal–occipital regions for high arousal, frequently highlighting F3/F4, T7/T8, and Pz/Oz—patterns that align with affective neurophysiology and support principled channel selection for low-density wearables. The resulting pipeline is reproducible, trains within minutes per fold on a single GPU, and offers an interpretable, deployment-friendly pathway to real-time affect monitoring.