综合信息系统研究中心知识库

随着脑科学与智能算法领域的不断发展，脑电研究领域与机器学习等研究领域具有越来越多的交集。一方面，对人脑生理结构和思维模式的研究能够促进智能算法的发展，另一方面，机器学习等智能算法也在逐渐扩展我们对于人脑的认识。近年来，越来越多的研究者开始采用机器学习的手段分析脑电信号，例如抑郁症、精神分裂症等脑疾病患者的脑电分析；事件相关电位、运动想象电位等脑-机接口算法；情绪、注意力等脑状态分析等。然而时至今日，基于脑电信号来分析人格特质的相关研究仍处于起步阶段，由于脑电数据中与个人特质相关因素并不像脑疾病、认知状态的特征那样显著，因此如何使用脑电信号自动识别被试者的人格特质水平一直是一项充满挑战的难题。组织承诺作为一项重要的人格特质，在许多职业群体或特殊人群中受到了强烈的关注。而传统的组织承诺评估是基于心理学量表来开展的，由于多种社会因素的影响，在实际应用场景中难以获得真实的结果。因此，本文探究了利用脑电生理信号来识别组织承诺水平的方法。
针对这一任务，本研究首次构建了组织承诺水平关联的静息态脑电数据集。本文采集了被试者的静息态脑电数据并使用组织承诺量表对被试者赋予组织承诺水平标签（高/低），通过提取多通道静息态脑电信号中的一系列不同类型的特征（包括功率谱分析特征、微状态分析特征、功能脑网络特征以及非线性动力学特征）、借助机器学习方法（支持向量机、逻辑回归、梯度提升决策树、K近邻算法、高斯朴素贝叶斯模型）来实现了组织承诺水平的二分类任务。本文在交叉验证实验中，评估了不同特征、不同分类器的准确率、召回率、精确率以及F1得分等不同指标，最终总结了在该任务下表现相对良好的静息态脑电特征（alpha波段功率谱密度、脑网络加权聚类系数、排序熵以及近似熵）。
在此基础之上，本文对静息态脑电分类实验的关键细节问题做了更进一步的研究。通过对比实验，验证发现睁眼状态脑电不适用于组织承诺水平的有效分类；同时，随着短时脑电样本的时间窗长度减小，分类器的各项指标呈现下降趋势；此外，本研究对比了不同脑区电极的分类性能差异，同时发现在只使用少量典型脑区的电极时，机器学习分类的各项指标均有所下降。为了进一步提升组织承诺水平分类的准确率等各项指标，本研究构建了基于Stacking（堆叠）策略的多层融合模型框架，对4秒时长的静息态脑电样本特征实现了82.6%的分类准确率以及0.827的F1得分；此外，为了克服脑电样本时长较小时的分类困难，本研究基于深度学习方法构建了一系列用于组织承诺水平分类的模型，包括基于时域脑电的一维卷积网络框架（EEG-1DCNN）、基于时域脑电的循环神经网络框架（EEG-LSTM）、基于脑电频谱图像的二维卷积网络框架（ESI-2DCNN），同时，考虑到脑电信号的空间拓扑性质，构建了一种基于脑电特征拓扑图的图卷积神经网络框架（EEG-GCN）。不仅如此，本文针对EEG-GCN的拓扑图邻接矩阵的构建指标进行了细致对比，最终基于“成对相位一致性（PPC）”与空间测地距离（geodetic distance）来实现EEG-GCN的分类——对0.5秒时长的极短静息态脑电能够达到79.1%的准确率以及0.788的F1得分，优于其他形式的深度学习分类框架，实现了对较短脑电样本的高准确率识别。
综上，本文针对组织承诺水平分类这一目标，对静息态脑电的特征以及不同分类模型的构建进行了较为全面的探索，这为今后进一步研究基于脑电的组织承诺分析或者其他人格特质分析提供了理论和方法基础。

With the continuous development of the field of brain science and intelligent algorithms, the field of EEG research and the field of machine learning research have more and more intersections. On the one hand, research on the physiological structure and thinking mode of the human brain can promote the development of intelligent algorithms; on the other hand, intelligent algorithms such as machine learning are gradually expanding our understanding of the human brain. In recent years, more and more researchers have begun to use machine learning to analyze EEG signals, such as EEG analysis of patients with brain diseases such as depression and schizophrenia; algorithm research on different brain-computer interfaces such as event-related potentials and motor imagery potentials; brain state analysis such as emotion, attention, etc. However, research on analyzing personality traits based on EEG signals is still in its infancy. Since the factors related to personal traits in EEG data are not as significant as the characteristics of brain diseases and cognitive states, how to use EEG signals to automatically identify the level of subjects' personality traits has always been a challenging problem. Organizational commitment, an important personality trait, has received strong attention in many occupational groups or special populations. Traditional organizational commitment assessment is carried out based on psychological scales. Due to the influence of various social factors, it may be difficult to obtain real results in practical application scenarios. Therefore, in this paper, methods for identifying levels of organizational commitment using brain electrophysiological signals are explored.
For this task, this study is the first to construct a resting-state EEG dataset associated with levels of organizational commitment. In this paper, the resting-state EEG data of subjects was recorded, and the subjects were assigned category labels using organizational commitment scales. A series of different types of features (including features based on power spectrum analysis, micro-state features, functional brain network features and nonlinear dynamic features) in multi-channel resting-state EEG signals were extracted. Then, the binary classification task of organizational commitment levels is realized using different machine learning methods (support vector machines, logistic regression, gradient boosting decision trees, K-nearest neighbor algorithm, Gaussian naive Bayes classifier). In cross-validation experiments, this paper evaluates the performance of different features and different classifiers based on different metrics such as precision, recall, precision, and F1 score. Therefore, the resting-state EEG features that perform relatively well under this task are summarized, namely alpha-band power spectral density, weighted clustering coefficient of brain network, permutation entropy and approximate entropy.
On this basis, this paper further studies the key details of the resting-state EEG classification experiments. Through comparative experiments, it is verified that the EEG in the eyes-open state cannot achieve effective classifications of organizational commitment levels; at the same time, as the time window lengths of EEG samples decrease, the indicators of the classifier show a downward trend; in addition, the study compared the differences in the classification performance of electrodes in different brain regions, and found that when only a small number of electrodes in typical brain regions were used, all indicators of machine learning classification decreased. In order to further improve various indicators such as the overall accuracy of classifications for organizational commitment levels, a multi-layer fusion model framework based on the Stacking strategy was constructed, which achieved an overall classification accuracy of 82.6% and an F1 score of 0.827 for the 4-second resting-state EEG samples; in addition, in order to overcome the classification difficulty when the lengths of the EEG samples are small, this study constructed a series of models for the classification of organizational commitment level based on deep learning methods, including EEG-based one-dimensional convolutional network framework (EEG-1DCNN), EEG-based recurrent neural network framework (EEG-LSTM), two-dimensional convolutional network framework based on EEG spectrogram images (ESI-2DCNN). Furthermore, considering the spatial topological properties of EEG signals, a graph convolutional neural network framework (EEG-GCN) based on EEG-feature topological graphs was constructed. Not only that, this paper makes a detailed comparison of the construction indicators of the topological adjacency matrix of EEG-GCN, and finally realizes the classification of EEG-GCN based on pairwise phase consistency (PPC) and spatial geodetic distance, which achieved an overall accuracy of 79.1% and an F1 score of 0.788 for very short resting-state EEG samples with a duration of 0.5 seconds. This outperforms other forms of deep learning classification frameworks, enabling high-accuracy recognition of shorter EEG samples.
To sum up, this paper conducts a comprehensive exploration of the characteristics of resting-state EEG and the construction of different classification models for the goal of classification of organizational commitment levels. This provides a theoretical and methodological basis for further research on EEG-based organizational commitment analysis or other personality trait analysis.