摘要:建立邻接图上的批量边删除聚类算法通用框架，提出基于高斯平滑模型的批量边删除判定准则，定义了适于聚类的邻接图的一般性质，提出并证明在kNN图基础上引入随机因子构造的随机kNN图，可以增强顶点之间的局部连通性，使聚类结果不再强烈依赖于某条边或某些边的保留或删除.RkNNClus算法简洁高效，依赖参数少，无需指定类簇数目，模拟和真实数据上的实验均有证明.



Abstract:By generalizing batch edge-removal clustering algorithm, the clustering problem can be separated into the deterministic problem of edge-remove and the construction problem of adjacent graph. Firstly, in this paper, an edge-removal criterion is proposed according to the shifting under the local Gaussian smoothing of data objects. Secondly, the properties of adjacent graph suitable for clustering are studied, and a random kNN (RkNN) graph is suggested by introducing the random factor into the kNN graph. A proof is given to show that RkNN graph can lead to the enhancement of the local connectivity of graph and less dependency between clustering results and the removal of certain edges. The RkNNClus is simple and efficient without specifying the number of clusters. The experiments on synthetic datasets and real datasets demonstrate the effectiveness of the method.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/5327