Individual Similarity Guided Transfer Modeling for EEG-based Emotion Recognition

Intelligent recognition of electroencephalogram (EEG) signals has been an important means to recognize emotions. Traditional user-independent method, which treats each individual's EEG data as independent and identically distributed (i.i.d.) samples and ignores destruction on i.i.d. condition caused by individual differences, usually has lower generalization performance. Although user-dependent method could alleviate above-mentioned problem, it faces difficulty in collection of sufficient training EEG data for each individual. In order to construct user-dependent model merely based on a small amount of training EEG data, we incorporate transfer learning framework and propose a individual similarity guided transfer modeling method for EEG-based emotion recognition. We first measure the similarities between individuals using maximum mean discrepancy (MMD), then utilize pre-existing EEG data of similar individuals to assist construction of user-dependent model for the target individual using an instance-based transfer learning algorithm named TrAdaBoost. We compared this method with traditional user-independent and user-dependent methods on DEAP dataset. Experimental results showed that our method could transfer useful knowledge from other individuals for user-dependent emotion recognition, which achieved classification accuracies of 66.1% and 66.7% on arousal and valence dimentions, respectively.