Domain Adaptation for Cross-Subject Emotion Recognition by Subject Clustering

Abstract

The high inter-subject variability in emotional EEG activities has posed great challenges for practical EEG-based affective computing applications. The recently popular domain adaptation strategy seemed to be a promising technique for addressing this issue, by minimizing the discrepancy of EEG data from different subjects. The present study proposed and implemented an extended Domain Adaptation method by introducing Subject Clustering (DASC). By clustering subjects based on the similarity of their emotion-specific EEG activities, the DASC method could make a flexible use of the available source domain information towards an optimized target domain application. Using the publicly available EEG dataset of DEAP, the DASC method achieved an average accuracy of 73.9±13.5% and 68.8±11.2% for binary classifications of the high or low levels of valence and arousal. Comparison with the state-of-the-art performance as well as the ablation experiments suggest the proposed DASC method as an effective extension to the conventional domain adaptation methods for EEG-based emotion recognition.