Abstract <p>Electroencephalogram (EEG)-based emotion recognition has attracted increasing attention due to its objectivity and ability to directly reflect brain activity. However, significant inter-subject variability and the non-stationary nature of EEG signals hinder model generalization, particularly in cross-subject and cross-session scenarios. To address this, multi-source domain adaptation has been widely applied in EEG-based tasks, offering a promising solution. Yet, it still faces challenges in distribution alignment instability and negative transfer from unreliable source domains. To this end, we propose a novel framework, Multi-Source Self-Guided Domain Adaptation (MSSGDA) in this work. MSSGDA introduces an ensemble branch to aggregate information across sources and a self-guided prediction aggregation module to dynamically suppress unreliable domains during inference, enabling more robust representations for cross-domain EEG emotion recognition. Additionally, during model training, MSSGDA aligns both marginal and conditional distributions to reduce domain discrepancies, while leveraging prediction entropy to encourage more confident predictions. Experimental evaluations conducted on the SEED and SEED-IV datasets demonstrate that MSSGDA achieves average accuracies of 88.26% and 74.17% in cross-subject experiment, and 83.09% and 79.89% in cross-session experiment, respectively. These results demonstrate that the proposed method substantially improves accuracy and generalization on unseen data, thereby strengthening the application potential of EEG-based emotion recognition.</p> Graphical abstract <p></p>