Towards Practical Emotion Recognition: An Unsupervised Source-Free Approach for EEG Domain Adaptation
Abstract
EEG-based emotion recognition systems face significant challenges when deployed across different domains due to limited labeled data and individual signal variability. We introduce a source-free unsupervised domain adaptation technique that operates without requiring access to original training data — a key advantage for privacy-sensitive applications. Our framework minimizes prediction discrepancies on confident samples and enforces local consistency by promoting similar predictions among reliable neighbours. Evaluated on DEAP, SEED, and DREAMER benchmarks, the method achieves substantial improvements over existing approaches.
Type
Publication
IEEE Transactions on Affective Computing
Authors
Post-Doctoral Fellow
Recently completed his PhD with Dr. Khan. Expert in physiological signal analysis and unsupervised domain adaptation, with a focus on cross-dataset generalization for ECG and EEG-based applications.
Authors
Associate Professor & Lab Director
Associate Professor and Director of the Multimedia Research Laboratory at Toronto Metropolitan University. Research spans multimedia signal processing, machine learning, and AR/VR with applications in healthcare and mental health.