一种基于排序子空间的高维聚类算法及其可视化研究

Clustering by Ordering Density-Based Subspaces and Visualization

为了有效地发现数据聚簇 ,尤其是任意形状的聚簇 ,近年来提出了许多基于密度的聚类算法 ,如DBSCAN ,OP TICS ,DENCLUE ,CLIQUE等 提出了一个新的基于密度的聚类算法CODU(clusteringbyorderingdenseunit) ,基本思想是对单位子空间按密度排序 ,对每一个子空间 ,如果其密度大于周围邻居的密度则形成一个新的聚簇 由于子空间的数目远小于数据对象的数目 ,因此算法效率较高 同时 ,提出了一个新的数据可视化方法 ,将数据对象看做刺激光谱映射到三维空间 。

Finding clusters on the basis of density distribution is a traditional approach to discover clusters with arbitrary shape Some density based clustering algorithms such as DBSCAN, OPTICS, DENCLUE, CLIQUE, etc have been explored in recent researches A new approach is presented, which is based on the ordered subspaces to find clusters The key idea is to sort the subspaces according to their density, and set a new cluster if the subspace is larger than its neighbors Since the number of the subspaces is much less than that of the data, very large databases with high dimensional data sets can be processed with high efficiency A new method is also presented to project high dimensional data, and then some results of clustering with visualization are demonstrated