Cross-domain transferable discriminant dictionary based sparse representation approach for EEG emotion-level recognition

Emotion is usually caused by complex psychological and physiological changes triggered by external stimuli. It is considered one of the abilities to guide one's thinking and action. Emotion-level recognition realizes psychological perception and evaluation by collecting various data characteristics of individual behavior and physiological level, and reversely deriving complex physiological psychological mapping. Emotional Electroencephalogram (EEG) signals are not easy to be hidden and forged, and using them for emotion-level recognition has high application value. However, EEG data are often collected in complex environments and scenes, and there are strong time and individual differences in EEG signals. This study proposes a cross-domain transferable discriminant dictionary based sparse representation (CTDDSR) approach for EEG emotion-level recognition. CTDDSR utilizes subspace projection strategy to find a suitable projection subspace. In this subspace, CTDDSR learns the shared dictionary to construct a strong connection between the source and target domains in the framework of dictionary based sparse representation. Different from the traditional projection matrices are completely domain-specific, our projection matrix for each domain is the combination of the domain-specific component and domain-invariable component. The former component is used to retain the individual domain information. By maximizing the cross-domain reconstruction error, the latter component is used to exploit the discriminant knowledge across domains in the shared latent subspace. In addition, to obtain the discriminant dictionary in the subspace, CTDDSR introduces linear discriminant analysis (LDA) to ensure the minimum of the intra-class reconstruction error and the maximum of the inter-class reconstruction error on sparse coding coefficients within each domain. Cross-domain EEG emotion-level recognition experiments are performed on the real EEG emotion dataset and verified that CTDDSR has excellent recognition performance.