Multi-source Information-Shared Domain Adaptation for EEG Emotion Recognition

EEG-based affective computing aims to provide an objective method to evaluate people’s emotional states in human-computer interactions, but it suffers the dilemma of individual differences in EEG signals. The existing approaches usually extract domain-specific features, which ignore the commonness of subjects or treat all subjects as one source for transfer. This paper proposes a novel multi-source information-shared domain adaptation framework for cross-subject EEG emotion recognition. In the proposed framework, we assume that all EEG data share the same low-level features, the shared representations and private components are captured by the shared extractor and private extractors, respectively. Besides the maximum mean discrepancy and diff losses, we also propose the was-loss to align the private domains for the purpose of extracting tight shared domain and thus enhancing the domain adaptation ability of the network. Finally, we build the domain-specific classifiers and shared classifier in parallel and dynamically integrate their predictions by the similarity of marginal distributions among domains. The experimental results on the SEED and SEED-IV datasets demonstrate that our framework outperforms the state-of-the-art domain adaptation methods with accuracies of 88.1% and 73.8% on average.