MS-SSCTGN: a multi-scale self-supervised contrastive spatiotemporal network for EEG-based emotion recognition

Abstract

Graph convolutional networks (GCNs), owing to their capability to effectively process non-Euclidean data, have become one of the dominant approaches for decoding emotional states from electroencephalography (EEG) signals. However, current brain region division strategies exhibit limitations in information extraction, and existing GCN-based models still face several challenges, including limited receptive fields, over-smoothing in deep graph networks, and insufficient spatiotemporal feature integration. To address these challenges, this paper proposes a dual-branch architecture termed MS-SSCTGN. Moving beyond conventional non-overlapping anatomical partitioning, this study first introduces an anterior-posterior overlapping brain region division strategy that allocates electrodes along the central coronal line to both anterior and posterior subgraphs simultaneously, establishing them as cross-regional information interaction hubs. Within this topological framework, the model initially utilizes a self-supervised contrastive learning pre-training module to uncover the intrinsic invariant representations of the data, thereby improving robustness to input perturbations and providing optimized parameter initialization. Subsequently, building upon these pre-trained representations, the network employs multi-scale graph convolutions to hierarchically aggregate spatial features, which broadens the effective receptive field and helps alleviate the tendency toward over-smoothing. Concurrently, the temporal convolutional network branch utilizes causal convolutions to capture temporal dynamics, achieving deep complementarity of spatiotemporal features. The proposed MS-SSCTGN demonstrates effectiveness in capturing complex emotional patterns in the subject-dependent EEG emotion recognition task, achieving an average recognition accuracy of 88.83% on the SEED dataset and 80.13% on the SEED-IV dataset.