基于DBN和RF的跨被试情绪识别研究

Cross⁃subject emotion recognition based on DBN and RF

目前情绪识别的分类方法很多,但情绪分类模型多具有被试依赖性,基于SEED数据集探索了跨被试情绪识别模型.首先将所有被试的脑电(Electroencephalogram,EEG)数据合并为一个被试,共提取675个trial三类情绪(正性(positive)、中性(neutral)、负性(negative)情绪)的短时傅里叶变换(Short-Time Fourier Transform,STFT)、离散小波变换(Discrete Wavelet Transformation,DWT)特征,并使用ReliefF特征选择算法对特征进行权重排序.其次,从排序好的特征中选择600个trial作为训练集,剩余的作为测试集;然后将K最近邻(K-Nearest Neighbor,KNN)、二次判别分析法(Quadratic Discriminant Analysis,QDA)、支持向量机(Support Vector Machine,SVM)、随机森林(Random Forest,RF)、深度置信(信念)网络(Deep Belief Network,DBN)五种分类算法作为分类器,对比研究选出最优的分类框架.结果表明,五种分类器的平均分类精度分别为:KNN 69.21%±3.4%,QDA 52.17%±9.41%,SVM 78.41%±3.8%,RF 83.49%±2.6%,DBN 81.73%±2.22%,可见RF的分类效果最好.分别计算每个分类模型对负性、中性、正性情绪的分类准确率,结果如下:不同分类器对正性情绪的识别效果都比较好;KNN,QDA,SVM对负性和中性情绪的分类效果较差,准确率不高;DBN和RF对负性和中性情绪的识别率较高,能有效地进行情绪识别.以上研究可望为跨被试的情绪识别模型提供参考.

At present,there are many classification methods for emotion recognition,but most of them are subject dependent.This study explores cross-subject emotion recognition model based on SEED dataset.Firstly,the EEG(Electroencepha-logram)data of all subjects are combined into one subject,and the Short-Time Fourier Transform(STFT)and Discrete Wavelet Transform(DWT)of 675 trial emotions(positive,neutral and negative emotion)are extracted.Secondly,600 trials are selected from the sorted features as the training set,and the rest as the test set.Then,K-Nearest Neighbor(KNN),Quadratic Discriminant Analysis(QDA),Support Vector Machine(SVM),Random Forest(RF)and Deep Belief Network(DBN)are used to rank the weights of features.The experimental results show that the average classification accuracy of KNN is 69.21%±3.4%,of QDA is 52.17%±9.41%,of SVM is 78.41%±3.8%,of RF is 83.49%±2.6%and of DBN is 81.73%±2.22%,respectively.Obviously,RF has the best classification effect.Also,the classification accuracy of each classification model for positive,neutral and negative emotion is calculated,and each classifier has good recognition effect on positive emotion,while KNN,QDA and SVM are poor to recognize negative and neutral emotions and difficult to get high accuracy.However,DBN and RF can overcome this problem and have a high recognition rate of negative and neutral emotions,which can effectively identify different emotions.This study is expected to provide a reference for cross-subject emotion recognition model.