基于正交回归和特征加权的脑电情感特征选择方法

EEG emotional feature selection method based on orthogonal regression and feature weighting

颅内容积传导效应导致大量脑电特征之间具有高度相关性,而这些高度相关的脑电特征无法为情感识别提供额外的有用信息,并且会降低基于脑电信号的情感识别效率.为了去除冗余信息和挑选有判别力的脑电特征,本文提出了一种基于正交回归和特征加权的脑电情感特征选择方法.与传统特征选择方法相比,该方法利用正交回归在脑电特征映射空间中保留更多的判别信息,更加适合于非线性和非平稳脑电信号的分析处理.为了验证所提出方法的性能,我们采集了由视频诱发的多通道脑电情感数据,并将所提出方法与4种常用的脑电特征选择方法进行了比较.实验结果证明了本文所提出方法能有效降低脑电特征集内冗余信息,并挑选出具有判别力的脑电特征子集.此外,通过分析由该方法所挑选的脑电特征类型,我们发现中心频率特征是最具判别力的脑电情感特征.该发现将为未来脑电情感特征提取研究提供新的思路.

The volume conduction effects of the human head result in highly correlated information among most EEG features.These highly correlated EEG features cannot provide additional useful information for emotion recognition and may reduce efficiency.This paper proposes a novel EEG emotional feature selection method called feature selection with orthogonal regression(FSOR)to reduce redundant information and select discriminative EEG features.Compared to common feature selection approaches,FSOR can utilize orthogonal regression to keep more discriminative information in the projection subspace for nonlinear and non-stationary EEG signals.To demonstrate the performance of our approach,we collected multichannel EEG recordings for emotion recognition and compared FSOR with four classical EEG feature selection approaches.The experimental results confirmed that the FSOR method outperformed the others in removing redundant features from the original EEG features.Furthermore,we found that the frequency at maximum power spectral density is the most discriminative EEG emotional feature.This discovery will inspire future studies on EEG emotional feature extraction.