Reinforced Multi-Source Transformer-LSTM Domain Alignment Framework for Cross-Session EEG Emotion Generalization

Abstract

Recognition of emotions via Electroencephalogram (EEG) suffers extreme performance degradation when under cross-subject and cross-session condition as a result of domain shift owing to disparities in physiological variability and environmental noise. To overcome this weakness, Multi-Source Reinforced Selective Domain Adaptation (MS-RSDA) framework is suggested, which can classify emotions robustly. Vision transformer (ViT) is incorporated into the model to extract spatial and spectral features, and Long Short-Term Memory (LSTM) networks are utilized to model the dependencies in time. Multi-source adaptation has the advantage of transferring discriminative representations selectively across multiple source domains, whereas a reinforcement learning system goes on to dynamically balance source weighting to avoid negative transfer. The methods used to reduce inter-domain disparity include adversarial alignment and matching of statistical distribution. Benchmark EEG experimental evaluation proves to be more accurate and stable than traditional CNN and single-source adaptation methods. The framework also facilitates behavioral scoring that is based on emotions that can be applied in mental health monitoring, adaptive learning, and intelligent human-computer interactive systems.