Research On Modeling Method Of Transferring Similar Individual For EEG Emotion Recognition

Emotion recognition is the most fundamental and key part in the field of affective computing.It is the important research content in the fields of artificial intelligence and human-computer interaction.Physiological signals,especially Electroencephalogram(EGG)signals,can display the activities of central nervous system directly and it is objective,real and difficult disguise,which have become the main data source in emotion recognition.However,EEG signals vary in individuality due to differences in gender,age,race,and health status,which leads to differences in EEG signals in different individuals with the same emotional state.The traditional emotion recognition methods usually assume that different individuals have the same feature space and its distribution,and the individual is set as an independent and identically distributed sample to construct the emotion recognition model,while ignoring the difference in data distribution of different individual EEG signals.The generalization performance of these model is poor.On the other hand,constructing a user-dependent emotion recognition model for each individual can alleviate the impact of individual differences effectively,but it requires the individual to provide a large amount of training data when constructing the model,while it is difficult,time consuming and costly to collect data.In order to solve the problem of constructing user dependence model for a single individual,this paper proposes a transfer similar individual modeling method for EEG emotion recognition.The method assists in constructing the emotion recognition model for target individual by transferring similar individual data with similar to the target individual in the distribution.The method effective alleviates the problem of collecting large scale training data faced by user dependent model while considering the influence of individual differences on EEG signals.The main work and contributions of this paper include the following three aspects.1)In view of the insufficient training data of target individuals and the influence of individual differences of EEG signals on the emotion recognition model,this paper analyzes the distribution characteristics of EEG data of different individuals,and uses the maximum mean discrepancy(MMD)to evaluate the similarity of the EEG data distribution of different individuals.Then we select the individuals with similar distribution to target individual and use the data as auxiliary training data to construct the emotion recognition model for target individual,which alleviates the demand for the target individual training data.2)In order to utilize the auxiliary EEG data provided by similar individuals effectively,this paper uses the TrAdaBoost method based on the instance transfer learning framework to transfer the auxiliary training data provided by similar individuals by increasing/reducing the misclassified weight of the target individual/similar individuals.We can construct an effective emotion recognition transfer model based on the MMD similar individual when the target individual provide a small amount of data.3)By comparing the performance of the method with the traditional emotion recognition model and the user dependence model on the DEAP(A Database for Emotion Analysis using Physiological Signals)data set,the method is superior to the traditional emotion recognition model when it has a small amount of target individual training data,and exceeds the classification effect of the user dependent model when it has the same target individual training data,which is verified on the Valence and Arousal the validity.At the same time,by comparing with the kernel mean matching(KMM)method,the advantages of the emotion recognition transfer model based on MMD similarity in the similar subject selection and training data weight update are proved.Finally,through in-depth exploration of the method,the research analyzes the influence of the number of similar individuals on the final classification results.The experiment proves that a small number of similar subjects is effective on constructing the emotion classification model,and too many individuals will lead to “negative transfer” phenomenon and affect the reliability of the model.In summary,the similar individual transfer modeling method for EEG emotion recognition can construct a reliable emotion recognition model for individual training data,which alleviates the influence from the individual differences of EEG signals effectively.And it provides a new way of thinking for emotion recognition modeling.