PDGCN: A progressive dual-branch graph convolution network for EEG emotion recognition

Electroencephalography (EEG) provides a direct measure of neural activity and offers an objective basis for emotion recognition. Existing graph neural network-based methods, however, often fail to capture both spatial topology and adaptive connectivity of EEG emotion signals, and typically focus on single-scale feature extraction. This study proposes a progressive dual-branch graph convolutional network (PDGCN) for EEG-based emotion recognition. The proposed framework jointly integrates progressive multi-scale temporal feature extraction with dual-branch graph convolution, enabling spatial modeling over temporally contextualized EEG representations. Specifically, PDGCN first extracts progressive multi-scale temporal features hierarchically from local to global scales. It then employs a dual-branch graph convolution strategy to extract and fuse spatial features: one branch encodes the intrinsic spatial topology of electrodes, while the other captures adaptive interactions between EEG channels by a learnable adjacency matrix. Cross-subject experiments on the SEED, SEED-IV, and DEAP datasets show that PDGCN achieves consistently competitive performance compared with representative existing methods. The proposed method offers an effective solution for EEG-based emotion recognition and introduces a new perspective for multi-scale and multi-view brain signal processing. Our codes are publicly available at https://github.com/wuminjin/PDGCN.