采用多任务特征融合的脑电情绪识别方法

Electroencephalogram emotion recognition method using multitask feature integration

特征选择与融合是提升脑电信号情绪解码精度的重要手段之一。然而,当前脑电情绪解码中的特征选择方法常忽略了脑电信号内在数据结构的隐含信息。该文提出一种基于近邻传播聚类的多任务特征融合方法,通过L_(2,1)范数约束实现稀疏特征选择,同时利用图拉普拉斯正则化保持不同子类间的潜在关系。该算法在不揭示真实样本标签的情况下,在子任务空间有效融合脑网络空间拓扑结构信息和微分熵信息,为高精度脑电信号情绪解码提供具有更高情绪表征能力的特征。DEAP和SEED数据集以及本实验室数据集的分析结果表明,该文提出的方法能显著提高脑电情绪解码的精度。

Feature selection and integration is one of the crucial approaches to improving the emotion decoding accuracy of electroencephalogram(EEG)signals.However,current methods often neglect the implicit information of the intrinsic data structure in EEG signals.Herein,a multitask feature integration method is proposed based on affinity propagation clustering.This method uses the L_(2,1)-norm constraint to select sparse features and uses graph Laplacian regularization to maintain potential relationships among different subclasses.In case of not disclosing real sample labels,the method has effectively integrated the spatial topology information of brain networks and differential entropy information in the subtask space,providing features with higher emotional characterization ability for the emotional decoding of high-accuracy EEG signals.The analytic results on DEAP and SEED datasets and the dataset of the laboratory show that the proposed method can markedly improve the decoding accuracy of EEG emotional decoding.