Research On Cross-subject EEG Emotion Recognition Method Based On Bi-mapping Joint Feature Transfer

EEG（Electroencephalogram）has been widely used for objective emotionrecognition because it is generated from the neural activities of central nervous system and is hard to camouflage.An obvious limitation is that the weak and non-stationary properties of EEG easily cause the individual differences in emotion recognition.This causes many machine learning methods to have poor generalization performance because the independent identically distributed assumption is no longer valid in cross-subject EEG data.To this end,transfer learning models were introduced to deal with this dilemma.However,existing models did not couple the feature adaptation process with the target label estimation process on one hand,and on the other hand they focused only on the recognition accuracy and have no sufficient investigation to the learned shared subspace.To solve the above issues,this thesis comes up with the following models in terms of feature transfer and EEG emotion activation patterns:1)A joint bi-projection domain adaptation and graph-based semi-supervised label estimation model for EEG emotion recognition（termed RAGE）.This model seamlessly unifies the three modules of domain adaptation,optimal graph learning and emotion state recognition together into a joint framework.Its internal coordination mechanism is:EEG feature migration realizes better data distribution alignment,and the gradually aligned data is convenient to better construct structured graph to describe the relationship between EEG data samples,further promoting the graph-based EEG emotion state estimation in the target domain;In turn,the emotional state estimation in the target domain effectively improves the accuracy of conditional distribution alignment and the effect of data distribution alignment.We evaluated the effectiveness of the proposed RAGE model on the benchmark SEED-IV emotional data set.The results showed that the two domain samples are effectively aligned in the subspace,and the performance of EEG emotion state recognition is significantly improved.Based on?_2,1mixed norm feature selection theory and the learning label specific features（LLSF）,we analyze the whole EEG emotion activation patterns and single EEG emotion activation patterns respectively.The research provided a reference for the study and analysis of EEG emotion activation patterns.2)An EEG feature transfer method combined with semi-supervised regression and bipartite graph label propagation（termed TSRBG）,which realizes the unified joint optimization of EEG feature distribution alignment and semi-supervised joint label estimation.The EEG emotional state in the target domain is estimated by a bi-model fusion strategy.The semi-supervised label propagation method based on sample-feature bipartite graph and semi-supervised regression method are combined to form a unified framework for joint common subspace optimization and emotion recognition.First,a sample-feature bipartite graph is constructed under the premise that samples with similar labels have similar feature distributions.This graph is used to distinguish the sample-feature connections between the source and the target domain for label propagation.Furthermore,a semi-supervised regression is used to learn a mapping matrix to describe the intra-domain connections between samples and labels,which aims to estimate the EEG emotional state of the target domain.We appraise the convincingness of the TSRBG model on the benchmark SEED-IV emotional data set.The results displayed that TSRBG has significantly better recognition performance in comparison with the advanced models.The EEG feature distribution differences between subjects are significantly minimized in the learned subspace.The key EEG frequency bands and channels for cross-subject EEG emotion recognition are analyzed by investigating the learned subspace,which provides more insights into the study of EEG emotion activation patterns.