基于无源域适应的脑电情绪识别

EEG Emotion Recognition Based on Source-Free Domain Adaptation

现有脑电情绪识别中的域适应方法利用源域数据及其特征分布来训练模型,不可避免地需要频繁访问源域,可能会导致源域受试者的隐私信息泄露。针对该问题,本研究提出一种基于高斯混合模型、核范数最大化和Tsallis熵的无源域适应(GNTSFDA)脑电情绪识别方法。首先,基于源域数据和本研究所提出的CNN和Transformer特征混合(CTFM)网络,利用交叉熵损失训练得到源域模型;然后,通过高斯混合模型聚类生成目标域数据的伪标签以构建分类损失;最后,基于伪标签和分类损失在目标域数据上对源域模型再训练以更新其参数,从而得到目标域模型,训练过程中还利用核范数最大化损失来提升模型预测的类判别性和多样性,同时利用Tsallis熵损失来减少模型预测的不确定性。GNTSFDA方法采用留一被试交叉验证的实验范式分别在SEED(源域14个受试者,目标域1个受试者)、SEED-IV(源域14个受试者,目标域1个受试者)和DEAP(源域31个受试者,目标域1个受试者)公开数据集上进行了实验。结果显示,在3个数据集上,目标域模型情绪识别的准确率分别为80.20%、61.20%和58.89%,相较于源域模型分别提升8.98%、7.72%和6.54%。GNTSFDA方法仅需要访问源域模型参数,而不是源域,从而有效地保护了源域受试者的隐私信息,在脑电情绪识别的实际应用中具有重要意义。

Existing domain adaptation methods in EEG emotion recognition utilize source domain data and feature distribution to train the model,which inevitably requires frequent access to the source domain and thus may lead to leakage of private information of the source domain subjects.To address this problem,this paper proposed a source-free domain adaptation EEG emotion recognition method based on the Gaussian mixture model,nuclear-norm maximization,and Tsallis entropy(GNTSFDA).First,based on the source domain data and the CNN and transformer feature mixture(CTFM)network,the source domain model was trained to obtain the source domain model using the cross-entropy loss.Then,the pseudo-labels of the target domain data were generated by clustering with the Gaussian mixture model to construct the classification loss.Finally,based on the pseudo-labels and the classification loss,the source domain model was re-trained on the target domain data to update its parameters to obtain the target domain model,and the nuclear-norm maximization loss was also utilized during the training process to enhance the class discriminative property and the diversity of the model predictions,and Tsallis entropy loss was utilized to reduce the model predictions'uncertainty.The GNTSFDA method was experimented on the SEED(14 subjects in the source domain,1 subject in the target domain),SEED-IV(14 subjects in the source domain,1 subject in the target domain),and DEAP(31 subjects in the source domain,1 subject in the target domain)public datasets,using a leave-one-subject cross-validation experimental paradigm.The results showed that on the three datasets,the accuracies of emotion recognition of the target domain model was 80.20%,61.20%,and 58.89%,respectively,which was an improvement of 8.98%,7.72%,and 6.54%,respectively,compared with that obtained from the source domain model.The GNTSFDA method only needs to access the source domain model parameters,instead of the source domain,therefore,effectively protected the privacy information of source domain subjects and is of great significance in the practical application of EEG-based emotion recognition.