DSGTANet: A semi-supervised dual-stream graph-temporal adaptation network for cross-subject continuous EEG emotion recognition.
TL;DR
Experimental results show that DSGTANet outperforms state-of-the-art methods in valence and arousal regression, achieving superior accuracy and robustness.
Abstract
Objective
Continuous emotion recognition using electroencephalogram (EEG) is a critical topic in brain-computer inter-faces (BCIs), with applications in mental disorder regulation and human-computer interaction. However, its practical appli-cation is significantly constrained by several key issues, including the scarcity of emotion labels, inter-subject distribution shifts, and heterogeneous regression noise. To address these issues, this study aims to develop a semi-supervised frame-work for accurate cross-subject continuous EEG emotion recognition. APPROACH We propose a novel Dual-Stream Graph-Temporal Adaptation Network (DSGTANet). First, we collect an EEG dataset of six dynamic emotional transitions with continuous labels from 40 participants. Then, DSGTANet is designed to consist of a spatial structural stream, a temporal dynamic stream, and a reliability-driven fusion and alignment module. Specifically, the spatial structural stream learns robust EEG channel-topology features through a learnable graph structure, structural augmentation, and a self-supervised consistency constraint. Meanwhile, the temporal dynamic stream employs a Transformer to capture long-range temporal dependencies in emotion evolution and introduces label-conditioned domain adversarial learning with an exponential mov-ing average (EMA) teacher to enhance pseudo-labeling for unlabeled samples. Furthermore, the reliability-driven module explicitly estimates dual-stream predictive uncertainty via heteroscedastic regression and performs precision-weighted product-of-experts fusion along with reliability-weighted conditional distribution alignment. MAIN
Results
Experimental results on our constructed dataset and the public DEAP dataset show that DSGTANet outperforms state-of-the-art methods in valence and arousal regression, achieving superior accuracy and robustness.
Significance
These findings highlight the considerable potential of the proposed DSGTANet for real-world EEG-based emotion decoding under conditions of limited labeled data and substantial subject variability.