一种基于局部线性嵌入的多流形学习算法

针对局部线性嵌入算法在处理多流形数据时失效问题,提出一种新的基于局部线性嵌入的多流形学习算法.采用cam分布寻找数据点的近邻,避免了近邻选取方向的缺失;同时在获取重建权值矩阵的过程中引入一个正则项约束,从而降低了算法对噪声的敏感度.通过对分布在不同流形上的高维数据实验后发现改进算法具有很好的降维效果.为了进一步验证算法的有效性,将改进后的算法对COIL-20数据库进行图像检索,结果表明该算法不仅有较好的降维效果而且在多类别多形状流形学习中有很好的实用价值.

A Multi-manifold Learning Algorithm Based on Locally Linear Embedding

In order to improve the correctness of locally linear embedding caused by multi-manifold data,a novel multi-manifold learning algorithm based on locally linear embedding is proposed in this paper.It accords the cam distribution to find neighbors of data points,and avoid the lack of the direction of neighbor selection;and a positive regularization is added in the gaining of the optimal reconstruction weight matrix to make it insensitive to the noise.Located in different manifolds for the high-dimensional data on the experiment to test the improved algorithm has a good effort of reducing dimension.Image retrieval using the COIL-20 database show that the improved algorithm has the practical value in the multi-class and multi-shape manifold learning.