Proceedings of the Annual Meeting of the Cognitive Science Society

Accurate assessment of working memory load (WML) is important for cognitive neuroscience and brain-computer interfaces, as WML reflects the recruitment of attentional control and memory resources during encoding, maintenance, and updating. However, WML-related EEG is highly nonstationary: ERP signatures vary in latency and amplitude, and these variations are amplified across subjects, leading to distribution shifts. We propose CMDA-GCN, an adaptive spatiotemporal graph convolutional network based on Confidence-Margin Semi-Supervised Domain Adaptation (CMDA). The model combines multi-scale temporal convolutions with time attention to capture transient ERP dynamics and accommodate latency variability, and uses adaptive graph convolutions to learn task-dependent inter-electrode interactions. To mitigate domain drift, CMDA-GCN integrates multi-source adversarial alignment with confidence-based pseudo-label learning and a hard-sample margin constraint, jointly suppressing pseudo-label noise and sharpening decision boundaries. Experiments on two datasets show consistent performance gains, and visualizations highlight workload-sensitive time windows and distributed interactions, supporting robust and interpretable WML estimation.