基于图划分的谱聚类方法的研究

Research of spectral clustering based on graph partition

谱聚类算法已得到机器学习领域的广泛关注,其算法思想来源于谱图理论,通过矩阵的特征分解获得数据的低维嵌入,并用于后续聚类中。介绍了谱聚类方法的基本原理和算法思想,指出现有的谱聚类算法中存在初始化敏感、如何自动确定聚类分组数以及如何降低问题复杂度等问题,并针对存在的问题提出了相应的解决方法。

Spectral clustering is newly concerned in machine learning.Spectral clustering algorithms derive from spectral graph theory, and solve eigenvalue decomposition of matrix to get the low dimensional embedding of data for later clustering.The spectral graph theory and algorithms framework are introduced,and some key problems of existing spectral clustering algorithms are pointed out,such as initialization sensitivity,how to automatically determine the clustering number and how to reduce the complexity of the problemetc.Lastly, the newly presented algorithms are highlighted.