中文摘要：

等距特征映射（Isomap）是一种新颖、高效的非线性降维技术,它的一个突出优点是只有两个参数需要设定,即邻域参数和嵌入维数。我们提出了一种新的估计Isomap的最优嵌入维数的算法,该算法使用执行Isomap过程本身所产生的数据来估计流形的最优嵌入维数,同时能确定邻域参数的最优值。通过与常用的残差估计方法的实例对比,说明这种算法对人造数据集和真实数据集都很有效,而且能更加合理、更加客观地估计出流形的最优嵌入维数。

英文摘要：

The isometric feature mapping （Isomap） algorithm is a novel and powerful technique for nonlinear dimensionality reduction, one of hers prominent advantages is only two parameters need to be set, i.e. the parameter of neighborhood and the embedding dimension. A new algorithm for estimating the optimal embedding dimension of Isomap was proposed. It used the data produced by performing Isomap itself to estimate the manifold’s optimal embedding dimension and it could also obtain the optimal value of the parameter of neighborhood. Experiments show that this algorithm is more reasonable and objective than residual variance technique and is effective to both artificial data and real-word data.