一种改进的扩散映射算法  

An improved diffusion maps algorithm

作　　者：徐丽丽[1] 闫德勤[2] 刘彩凤[2] 贾洪哲[2] 

机构地区：[1]辽宁师范大学数学学院,辽宁大连116029 [2]辽宁师范大学计算机与信息技术学院,辽宁大连116029

出　　处：《微型机与应用》2015年第8期23-25,28,共4页Microcomputer & Its Applications

摘　　要：扩散映射(Diffusion Maps)是一种基于流形学习的非线性降维方法。基于对扩散映射的研究,提出了一种新的非线性降维算法。根据近邻点分布的不同和模糊聚类原理,新算法定义了扩散映射算法构建权值矩阵的误差近似系数,并采用改进的距离公式来选取样本点的近邻点,很大程度地降低了近邻点的选取对降维效果的影响。实验结果表明,新算法有效地保持了高维数据中的流形结构,具有更好的降维效果,并在基于内容的图像检索中达到很高的查准率,新算法的有效性和优越性得到了证实。Diffusion Maps based on manifold learning is a non-linear dimensionality reduction method. Based on the study of Diffusion Maps algorithm, a new dimensionality reduction algorithm is proposed. According to the distribution of the neighboring points and fuzzy clustering theory, the new algorithm defines the error of approximately coefficient of constructing weights matrix, and it makes use of improved distance formula to select the neighboring points of sample points. To a large extent, it can reduce the influence of the selection of neighbors for dimension reduction. The experimental results show that new algorithm can effectively keep the manifold structure of high dimension data and obtain better dimension reduction results, and it has high retrieval precision in the image retrieval based on content. Validity and superiority of the new algorithm are confirmed.

关 键 词：扩散映射 降维 流形学习 聚类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]