Cross-subject electroencephalogram(EEG)emotion recognition suffers from severe domain shift and negative transfer.A dynamic region-aware adversarial graph convolutional network(DRA-GCN)is proposed.A lightweight region-aware graph convolutional network(GCN)is adopted to extract brain topological features.A low-intensity gradient reversal adversarial mechanism combined with an orthogonality constraint is used to disentangle emotional features from identity noise.A domain-discrimination gated mechanism is utilized to dynamically weight multi-source experts

so as to obtain personalized predictions for target subjects.On SEED and SEED-IV

the average accuracy reaches 90.34% and 78.15%

the standard deviation is reduced by approximately 52% compared with strong alignment methods

and the parameter count is only 0.52M.