一种基于高斯隐变量模型的分类算法

摘要

高斯过程隐变量模型是近年来新兴的无监督降维方法,它可以找到高维数据的低维流形结构.但是由于高斯过程隐变量模型是无监督的概率降维方法,所以当数据集中的样本有类别标记信息时,高斯过程隐变量模型不能利用这些监督信息,实现分类的任务.为了使高斯过程隐变量模型可以处理分类任务,文中提出了一种监督的高斯过程隐变量模型分类模型.通过最大化后验似然的方法确定观测数据在隐空间的坐标,同时可以完成分类任务.实验结果证明了该模型可以有效地用于分类.%Gaussian process latent variable model is a new probabilistic approach for dimensionality reduction. It can obtain a low-dimensional manifold of a data set in an entirely unsupervised way. However, when there is some supervised information in the data set, Gaussian process latent variable model cannot use this information for supervised tasks, e. G. , classification and regression learning. For this purpose, a supervised Gaussian process latent variable model for classification is developed. The maximum-a-posterior algorithm is employed to estimate all latent variables position. Compared with the traditional Gaussian process latent variable model, the supervised version of this model shows more advantages in experiments.