Towards Practical Emotion Recognition: An Unsupervised Source-Free Approach for EEG Domain Adaptation

Source raw items (1)

• arXiv2026-07-23 07:15:13
  Towards Practical Emotion Recognition: An Unsupervised Source-Free Approach for EEG Domain Adaptation

  Emotion recognition is crucial for advancing mental health, healthcare, and technologies such as brain-computer interfaces. EEG-based models, however, struggle in cross-domain settings due to the high cost of labeled data and signal variability across individuals and recording conditions. Unsupervised domain adaptation typically requires access to source data, which is often infeasible because of privacy and computational constraints. Source-free unsupervised domain adaptation (SF-UDA) removes this requirement, but it has not yet been applied to emotion recognition. We propose an SF-UDA approach for cross-domain EEG emotion classification, built on a multi-stage framework that adapts to the target domain without source data. Dual-Loss Adaptive Regularization (DLAR) minimizes prediction discrepancies on confident samples and aligns predictions with expected pseudo-labels. Localized Consistency Learning (LCL) enforces local consistency by promoting similar predictions among reliable neighbors. Together, these components address domain shift and reduce the impact of noisy pseudo-labels, a key challenge in SF-UDA. Experiments on DEAP, SEED, and DREAMER show that our method significantly outperforms state-of-the-art approaches, reaching 65.84% and 58.87% accuracy on SEED and DREAMER when trained on DEAP, and 58.99% and 67.08% on DEAP and DREAMER when trained on SEED. It detects both positive and negative emotions well, making it suitable for practical emotion recognition applications. Code available at: https://github.com/RyersonMultimediaLab/EmotionRecognitionSF-UDA