一种鲁棒的概率核主成分分析模型

A Robust Probabilistic Kernel Principal Component Analysis Model

摘要:

大数据时代面临的数据维数越来越高,对数据降维处理越发显得重要。经典的主成分分析模型已被证明是一种有效的数据降维方法。但它在处理非线性、存在噪声和异常点的数据时存在效果较差的问题。对此,文章提出了一种鲁棒概率核主成分分析模型。该模型将核方法与基于高斯隐变量模型的极大似然框架相结合,用多元 t分布作为先验分布,以同时解决主成分分析在这 3个方面的弊端。提出混合鲁棒概率核主成分分析模型,使其可直接用于对混合的非线性数据进行降维和聚类分析。在不同数据集上进行的实验结果表明,与标准的混合概率核主成分分析模型相比,文中模型在数据聚类方面有更高的准确率。

Abstract:

The dimension of the processed data have become more and more higher, so dimensionality reduction becomesmore and more important. The classical PCA (Principal component analysis) has proven to be an effective dimensionalityreduction method. But its effect was poor when used it to disposing nonlinear, noise and outliers data set, so, a robust prob.abilistic kernel principal component analysis model (RPKPCA) was proposed. It combined kernel method with maximumlikelihood frame based on Gaussian process latent variable model and used t-distribution as prior distribution to solve itsthree disadvantages at the same time. In addition, a mixtures of robust probabilistic kernel principal component analysismodel (MRPKPCA), and it could be used directly to dimensions reduction and data mining of mixture and nonlinear data.The experimental results in different data set showed that the model of proposed in this paper had higher accuracy than thestandard probabilistic kernel principal component analysis model.