摘要： 针对传统谱聚类算法中亲合矩阵构造不准确和聚类结果不稳定的问题，提出一种基于邻里关系传播与模式合并的谱聚类算法。根据邻里关系传播原则更新子集内样本的相似度，设计局部最大相似值更新方法更新子集间样本的相似度，使用模式合并技术对子集个数较多的集合加以合并得出粗类，再对粗类间样本相似度进行二次更新，构造出亲合矩阵并将其用于谱聚类运算。实验结果表明，二次更新后，同类中样本的相似度被相对性放大，而不同类中样本的相似度则相对性缩小。与近邻传播的谱聚类算法相比，使用该算法能够得到更准确、稳定的聚类结果。

关键词: 谱聚类, 邻里关系传播, 亲合矩阵, 模式合并, 相似度, 二次更新

Abstract: Aiming at the problem that affinity matrix is constructed inaccurately and the clustering result is unstable in traditional spectral clustering algorithm, a spectral clustering algorithm based on neighborly relation propagation and mode merging is proposed in this paper. According to the principle of neighborly relation propagation, first update the similarity between samples in same subset, then it designs a local-max similarity updating method to update the similarity between samples in different subsets, uses the mode merging technology to merge these subsets whose numbers are more than the real clustering’s to obtain the coursing cluster, and further to update the similarity between samples in different coursing cluster and achieve the final affinity matrix. It applies this matrix to realize the spectral clustering. Experimental results show that after the secondary updating, the similarity between samples in the same cluster is relatively enlarged, and the similarity between samples in the different clusters is relatively reduced. Compared with the neighbor propagation spectral clustering algorithm, using the proposed algorithm can obtain the more accurate and stable clustering results.

Key words: spectral clustering, neighborly relation propagation, affinity matrix, mode merging, similarity, secondary updating

中图分类号: 

• TP391.4