Three-Stage Semi-Supervised Multisource Adversarial Learning for EEG Signal Processing and Emotion Detection

Abstract

Electroencephalogram (EEG) signals exhibit immense potential in emotion recognition due to their high temporal resolution and direct measurement of brain activity. However, intersubject variability in EEG measurements leads to differences in distributions across participants, limiting the robustness and generalization capabilities of EEG-based emotion recognition systems. To address this challenge, this article proposes a three-stage semi-supervised multisource adversarial network (TS-SMAN) with limited labeled data. In the first stage, a frequency-domain graph convolutional network (GCN) based on brain region priors is constructed. Through supervised learning, spatial dependencies between brain regions are modeled to obtain preliminary signal feature representations. The second stage introduces feature consistency constraints based on exponential moving averages (EMAs), utilizing unlabeled source-domain EEG data to enhance the stability and robustness of feature representations. In the third stage, each source domain was treated as an independent domain. Feature alignment between source-domain and target-domain measurements was achieved through a global–local adversarial discriminator, maximum mean discrepancy (MMD) constraints, and dynamic adversarial weights. In cross-subject and cross-session experiments on the SEED and SEED-IV datasets, TS-SMAN achieves classification accuracies of 88.9 %/90.1 % (SEED) and 71.6 %/74.2 % (SEED-IV), which demonstrates the effectiveness of the proposed model in dealing with limited labeled samples.