Research On Emotion Recognition From EEG Signal Based On Metric Learning

With the rapid development of artificial intelligence and human-computer interaction,related research on emotional computing has received extensive attention.EEG records changes in the potential generated by brain neuron activity,and it is most closely related to human emotions,and therefore has important significance in the field of emotional recognition.The EEG signal emotion recognition depends on the similarity measure between the input EEG data samples,and the similarity measure is achieved by the sample distance measure.At present,the research on EEG signal emotion recognition mainly revolves around effective feature extraction and classification algorithm.It lacks the research on EEG data sample metric.This paper applies metric learning to EEG emotion recognition.First of all,aiming at avoiding the problem of incorrect emotion labe ling of EEG data due to human subjectivity,fatigue,concentration,and other factors,in order to avoid the influence of abnormal samples on the subsequent measurement models and the construction of classification models,this paper studies the abnormal sample processing method in EEG signal emotion recognition.The EF algorithm integrates multiple classifiers to detect and reject abnormal samples at once through voting.Based on the integration of multiple classifiers,the MEIF algorithm obtains a clean E EG sample training set through multiple filtering,anomaly score measurement,abnormal sample rejection,and multiple iterations.Finally,in the SEED standard emotion recognition EEG dataset,based on the SVM method,the accuracy rate of emotion recogniti on increased by 0.8% and 1.6%.Second,for the Euclidean distance measurement of EEG data samples,this paper studies the EEG signal emotion recognition method based on LMNN metric learning.By using the Mahalanobis distance measurement method,the original EEG sample is mapped to a new feature space using the Markov matrix,and the EEG characteristics that contribute differently to the emotion are discriminated,and then more accurately measure similarities between EEG data samples.In the end,compared to using SVM,the accuracy rate of emotion recognition increased by 5.3%.In addition,the abnormal sample was processed in conjunction with the second chapter,and the accuracy of emotion recognitio n was further increased by 1.2%.Finally,aiming at the problem of distribution of EEG data samples due to the nonstationarity of EEG signals,this paper studies EEG signal emotion recognition method based on MMD-ML metric learning.By learning the known information of the emotional lable in the source domain and the distribution information of the unknown samples in the target domain,an optimal feature mapping method is obtained to map the samples of the source and target domains to the high-dimensional RKHS space.In this space,the source and target domains have the smallest MMD distances,which makes the distribution of samples as identical as possible,thus solving the problem of different sample distributions in the source and target domains.Finally,compared to the SVM method,the accuracy rate of emotion recognition across individuals and across time increased by 16.7% and 20.8%,respectively.In addition,the abnormal sample was processed in conjunction with the second chapter,the accuracy of emotion recognition was further increased by 0.7% and 0.3%.