A Coarse-to-Fine Semisupervised Framework for Reliable Cross-Subject Emotion Measurement Using EEG Signals

Abstract

Electroencephalography (EEG)-based emotion recognition faces critical challenges due to label scarcity and intersubject variability. While semisupervised learning mitigates annotation costs, most existing domain adaptation methods rely on global alignment strategies that inadequately capture fine-grained categorical variations in cross-subject EEG signals. To address these limitations, we propose a coarse-to-fine semisupervised domain adaptation (CFSDA) framework that progressively aligns distributions at both domain and class levels. At the coarse-grained level, adversarial training aligns global distributions across labeled source, unlabeled source, and target domains. A domain-independent representation learning network (DRLN) based on graph convolutional networks with a shuttle attention (SA) mechanism captures domain-invariant spatiotemporal representations, modeling channelwise and spectralwise dependencies. At the fine-grained level, a prototype-guided contrastive learning (CL) mechanism leverages spectral–spatial prototypes to adaptively categorize target samples into low-consistency and high-consistency subsets based on prototype consistency, applying customized interdomain and intradomain contrastive objectives for class-level alignment. An online neighbor-aware pseudo-labeling strategy generates reliable pseudolabels through entropy-weighted reweighting. Extensive experiments demonstrate that CFSDA achieves state-of-the-art performance in label-scarce cross-subject emotion recognition.