Research On Local Linear Embedding Joint Knowledge Transfer Method For EEG Emotion Recognition

Emotion plays an important role in the communication and interaction.It enables people to express themselves beyond the field of language.With the more and more in-depth interaction between us and machines,emotion recognition has become the focus research in the field of Human-Machine Interaction(HMI).Emotion recognition helps to build a more friendly HMI system.Among the many physiological signals,EEG not only reflect the electrophysiological activities of brain nerve cells on the cerebral cortex or scalp surface,but also contain a large number of physiological and disease information.Moreover,EEG is widely used in emotion recognition because it is difficult to disguise,portable and non-invasive.However,human beings will be affected by external factors such as the growth environment and internal factors such as their own physical conditions,the EEG of different people will have certain differences.Individual differences will lead to significant differences in the distribution of EEG data among different subjects.Traditional emotion recognition models cannot be well shared across subjects.And training a personal model for new subjects requires collecting a large amount of labeled data.Therefore,in some practical applications,we hope to acquire the model as soon as possible,while reducing the demand on the amount of label data for new subjects.In order to achieve this goal,we proposed a joint knowledge transfer method based on locally linear embedding.The target domain is divided into labeled data for finding optimal parameters and unlabeled data for testing.In order to reduce the difference of EEG between the target domain and the source domain,we use a small amount of labeled data from new subjects to transfer,then we use the source model for subsequent emotion recognition.The main work and contributions of this paper are as follows:1.A Local Linear Embedding Enhanced Joint Knowledge Transfer(LLEJKT)method is proposed.In this method,a new representation of each source domain is obtained by finding two projection matrices,and the two projection matrices are expected to meet the following requirements: First,the source domain is still discriminable in the low-dimensional space after projection,i.e.the source domain is discriminable;Second,the target domain can still maintain the local linear relationship of the original space in the low-dimensional space,i.e.target domain local linear relationship preservation;Third,the data distribution difference between the source domain and the target domain in the low-dimensional space is minimized.Finally,the distance between the two subspaces is as small as possible.2.Homogeneous Local Linear Embedding Enhanced Joint Knowledge Transfer(HLLEJKT)is an improved method based on LLEJKT.Unlike LLEJKT,HLLEJKT takes into account the labeled information in the target domain.Similarly,we hope that the two projection matrices can meet the following requirements: Firstly,keeping the source domain discriminable;Secondly,the local linear relationship between the same kind of samples in the original space can be maintained in the target domain in the low-dimensional space,i.e.to maintain the local linear relationship of the same kind of data in the target domain.Thirdly,minimization of joint probability distribution shift between two domain.Finally,the distance between the two subspaces are smaller.In addition,we also give the representation of LLEJKT and HLLEJKT in nuclear space.3.In order to verify the effectiveness of LLEJKT and HLLEJKT,this paper designed three experiments on benchmark SEED and benchmark SEED-IV:single-source to single-target(STS)transfers?multi-source to single-target(MTS)transfers and single-source contains few samples to single-target(SFSTS).Experimental results show that our method can alleviate individual differences to a certain extent.Compared with other transfer learning methods and non-transfer learning methods,our method has stronger applicability.In summary,the LLEJKT method and HLLEJKT method can dig deeply into the data distribution information of the source domain and the target domain,then learn a new feature space that is conducive to emotion recognition.Both methods have good robustness and provide a new idea for the adaptive problem of emotion recognition based on EEG in multi-subject domains.