有监督的噪音流形学习算法

Algorithm of Supervised Learning on Outlier Manifold

流形学习算法是维度约简与数据可视化领域的重要工具,提高算法的效率与健壮性对其实际应用有积极意义。经典的流形学习算法普遍的对噪音点较为敏感,现有的改进算法尚存在不足。本文提出一种基于监督学习与核函数的健壮流形学习算法,把核方法与监督学习引入降维过程,利用已知标签数据信息与核函数特性,使得同类样本变得紧密,不同类样本变成分散,提高后续分类任务的效果,降低算法对流形上噪音的敏感性。在UC I数据与白血病拉曼光谱数据上的实验表明本文改进的算法具有更高的抗噪性。

Manifold learning algorithm is an important tool in the field of dimension reduction and data visualization.Improving the algorithm＇s efficiency and robustness is of positive significance to its practical application.Classical manifold learning algorithm is sensitive to noise points,and its improved algorithms have been imperfect.This paper presents a robust manifold learning algorithm based on supervised learning and kernel function.It introduces nuclear methods and supervised learning into the dimensionality reduction,and takes full advantage of the label of some data and the property of kernel function.The proposed algorithm can make close and same types of samples and distribute different types of samples,thus to improves the effect of the classification task and reduce the noise sensitivity of outliers on manifold.The experiments on the UCI data and Raman data of leukemia reveal that the algorithm has better noise immunity.