基于谱特征嵌入的脑网络状态观测矩阵降维方法

发布时间：2018-01-30 21:49

本文关键词： 高维数据降维功能脑网络脑网络状态观测矩阵谱特征嵌入算法动态特性　出处：《计算机应用》2017年08期 　论文类型：期刊论文

【摘要】：针对基于功能核磁共振(f MRI)重构的脑网络状态观测矩阵维数过高且无特征表现的问题,提出一种基于谱特征嵌入(Spectral Embedding)的降维方法。该方法首先计算样本间相似性度量并构造拉普拉斯矩阵;然后对拉普拉斯矩阵进行特征分解,选取前两个主要的特征向量构建2维特征向量空间以达到数据集由高维向低维映射(降维)的目的。应用该方法对脑网络状态观测矩阵进行降维并可视化在二维空间平面,通过量化类别有效性指标对可视化结果进行评价。实验结果表明,与主成分分析(PCA)、局部线性嵌入(LLE)、等距映射(Isomap)等降维算法相比,使用该方法得到的脑网络状态观测矩阵低维空间的映射点有明显的类别意义表现,且在类别有效性指标上与多维尺度分析(MDS)和t分布随机邻域嵌入(t-SNE)降维算法相比,同一类样本间平均距离Di指数分别降低了87.1%和65.2%,不同类样本间平均距离Do指数分别提高了351.3%和25.5%;在多个样本上的降维可视化结果均有一定的规律性体现,该方法的有效性和普适性得以验证。
[Abstract]:In view of the problem that the dimension of state observation matrix of brain network based on functional nuclear magnetic resonance (MRI) reconstruction is too high and there is no characteristic representation. A dimensionality reduction method based on spectral feature embedding (Spectral embedding) is proposed, in which the similarity measure between samples is first calculated and Laplace matrix is constructed. Then the characteristic decomposition of Laplace matrix is carried out. Selecting the first two main Eigenvectors to construct the 2-D Eigenvector Space to achieve the data set Mapping from High to low Dimension (dimensionality reduction). This method is used to reduce the dimension of the state observation matrix of the brain network and visualize in the two-dimensional space plane. The experimental results show that, with the principal component analysis (PCA), local linear embedding (LLEs). Compared with the isometric mapping isomap algorithm, the mapping points in the low-dimensional space of the state observation matrix of the brain network obtained by this method have obvious category significance. And compared with MDS (Multidimensional scale Analysis) and t-SNEs (random neighborhood embedding) algorithm in class validity. The average distance Di index of the same kind of samples decreased by 87.1% and 65.2, and the average distance do index of different samples increased by 351.3% and 25.5, respectively. The dimensionality reduction visualization results on several samples are regular, and the effectiveness and universality of the method are verified.
【作者单位】：昆明理工大学信息工程与自动化学院;
【基金】：国家自然科学基金资助项目(61263017) 昆明理工大学教学改革项目(10968397)~~
【分类号】：O157.5;R445.2
【正文快照】： 0引言人类大脑由上千亿个神经细胞组成,相当于银河系的星体总数,它长期以来保持着神秘色彩,人们对其知之甚少。近些年世界上各个国家纷纷启动人类大脑计划或与人类大脑相关的研究计划,可见研究人类大脑的重要性。如今功能核磁共振成像(functional Magnetic Resonance Imaging,

【参考文献】