Proceedings of the Annual Meeting of the Cognitive Science Society

Reliable decoding of cognitive fatigue from Electroencephalography (EEG) is essential for monitoring sustained attention but is hampered by significant inter-subject variability in neural dynamics. To address the resulting domain shift, we propose the Multi-Level Domain Adaptation Network with an Enhanced Spatio-Temporal Graph (MLDA-STG). Our approach uniquely models the evolving functional connectivity of the brain during cognitive decline through a dynamic graph fusion network, while a collaborative learning mechanism disentangles personalized neural traits from shared fatigue patterns. Crucially, we introduce a Multi-Level Alignment framework that goes beyond global adaptation to align conditional distributions, preserving the semantic structure of cognitive states across individuals. Evaluations on the SEED-VIG and SADT datasets demonstrate that MLDA-STG achieves state-of-the-art performance, offering a robust solution for calibration-free cognitive state assessment. The source code is available at: https://github.com/zjh-sys/MLDA-STG.git.