Crosssubject distribution shift limits the generalization ability of electroencephalogram(EEG)-based emotion recognition models.To address this issue

a domain selection and contrastive learning-based multi-source domain adaptation(DSCL-MSDA)framework is proposed.The method constructs a cost matrix based on electrode position and frequency band information

selects the most relevant source domains using Sinkhorn distance

mitigates negative transfer

and enhances sample diversity through intra-class augmentation.In feature modeling

DSCL-MSDA employs shared and domain-specific feature extraction mechanisms

and combines a multihead discriminator to generate consensus pseudo-labels.Stable training is ensured by jointly optimizing classification

contrastive learning and other loss functions

and introducing a staged gain strategy.Experimental results show that DSCL-MSDA outperforms existing methods in accuracy and robustness on the SEED and SEED-IV datasets

which demonstrates its strong capability for cross-domain adaptation.