Electroencephalographic (EEG) signals are a critical indicator for recognizing emotional states, and employing EEG-based emotion recognition constitutes a key component in the diagnosis and rehabilitation of psychiatric disorders. Most current research in the field of emotion recognition focuses on the spatial–temporal information of EEG signals and does not adequately account for the variability of EEG signals across different individuals (cross‑subject) and across different time periods (cross‑session) . Domain adaptation methods effectively address these kinds of problems. The paper proposes a multi‑source domain‑adaptation method for EEG‑based emotion recognition. It transfers each source domain separately to the target domain, thereby avoiding interference from multiple EEG source domains. First, introduce a linear‑model–based shared feature extractor to obtain low‑level domain‑invariant features. Next, incorporate a multi‑head self‑attention–based domain‑specific feature extractor to capture domain‑specific characteristics. Through a feature processing mechanism, progressively eliminate domain‑specific interference factors such as physiological differences and emotional fluctuations. Finally, employ a long short‑term memory–based domain‑specific classifier to capture the temporal dependencies in the input data, enhancing the model’s ability to represent complex features. Conduct experiments on the SEED and SEED‑IV datasets. For cross‑subject evaluation, the accuracies are 90.02% and 91.57%, respectively; for cross‑session evaluation, the accuracies are 67.09% and 69.27%, respectively. The experiments show that our proposed method outperforms several baseline models in both cross‑subject and cross‑session emotion recognition tasks, fully validating the model’s effectiveness and generalization capability.