摘要
针对脑电信号(electroencephalogram,EEG)情感识别中因标签缺失导致模型泛化性不足以及单任务自监督学习易过拟合的问题,提出了一种基于EEG频域特征的图神经网络模型,采用自监督多任务学习框架进行表征学习和情感识别。提取EEG数据的微分熵特征并构建图结构表示,通过通道掩蔽、频率掩蔽、空间拼图和频率拼图四种自监督任务进行多任务学习。结合切比雪夫图神经网络提取深层特征,通道掩蔽和频率掩蔽任务通过重建模块计算损失,空间拼图和频率拼图任务通过分类模块计算损失。训练完成后,冻结特征提取器参数并用于下游情感识别任务。在SEED和DEAP数据集上的实验结果显示,依赖被试的情感分类准确率在SEED数据集上达到89.87%(三分类),在DEAP数据集上,唤醒度和效价维度的两分类准确率分别为88.03%和89.70%;而独立被试的准确率分别在SEED数据集上为72.03%,在DEAP数据集上为65.38%和61.29%。这些结果表明,所提方法有效提升了分类性能,缓解了过拟合问题,且优于现有方法。
To address the issues of the insufficient model generalization due to label scarcity and the overfitting tendency of single-task self-supervised learning in electroencephalogram(EEG)emotion recognition,this paper proposes a graph neural network model based on EEG frequency-domain features,within a self-supervised multi-task learning framework to enhance representation learning and emotion recognition.Differential entropy features are extracted from EEG data and represented as graph structures.Multi-task learning is performed through four self-supervised tasks:channel masking,frequency masking,spatial jigsaw,and frequency jigsaw.A Chebyshev graph neural network is utilized to extract deep-level features,where the channel masking and frequency masking tasks compute losses via a reconstruction module,while the spatial jigsaw and frequency jigsaw tasks compute losses via a classification module.After training,the feature extractor parameters are frozen and applied to downstream emotion recognition tasks.Experimental results on the SEED and DEAP datasets demonstrate that the subject-dependent emotion classification accuracy reaches 89.87%(three-class)on the SEED dataset,while on the DEAP dataset,the binary classification accuracies for the arousal and valence dimensions are 88.03% and 89.70%,respectively.For subject-independent scenarios,the accuracies are 72.03% on the SEED dataset and 65.38% and 61.29% on the DEAP dataset.These results indicate that the proposed method effectively improves classification performance,mitigates overfitting,and outperforms existing approaches.
作者
陈景霞
李小池
王倩
张鹏伟
CHEN Jingxia;LI Xiaochi;WANG Qian;ZHANG Pengwei(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi’an 710021,China;School of Frontier Science and Technology Transfer,Shaanxi University of Science and Technology,Xi’an 710021,China)
出处
《计算机工程与应用》
北大核心
2025年第22期205-214,共10页
Computer Engineering and Applications
基金
国家自然科学基金(61806118)
陕西科技大学博士科研启动基金(2020BJ-30)。
关键词
脑电信号
情感识别
图神经网络
自监督学习
多任务
electroencephalogram
emotion recognition
graph neural network
self-supervised learning
multi-task
作者简介
通信作者:陈景霞(1979-),女,博士,教授,CCF会员,研究方向为脑电信号与脑机接口、情感识别,E-mail:chenjingxia@sust.edu.cn;李小池(2001-),男,硕士研究生,研究方向为脑电情感识别、自监督;王倩(2002-),女,硕士研究生,研究方向为对比学习、脑电情感识别;张鹏伟(1977-),男,硕士,教授,研究方向为脑电信号处理与脑机接口应用。
