一种改进相似度参数估计的半监督谱聚类算法

Improved Similarity Parameter Estimation for Semi-supervised Spectral Clustering Algorithm

谱聚类算法的相似度参数对聚类效果有着至关重要的影响。本文将启发式思想引入到相似度计算中,通过对距离矩阵的搜索,找到数据的合理分界点,并利用它得到相似度计算时所需的参数。同时利用成对限制先验信息引导聚类过程,从而提高聚类效果。数据实验验证本文所提方法是可行的,并且具有很好的聚类效果。

Similarity parameter of spectral clustering algorithm plays an important role in clustering effectiveness. This paper introduces a heuristic technique into a spectral clustering algorithm to learn similarity parameter. The reasonable boundary point can be found by searching distance matrix. The pairwise constrains prior information is used to conduct clustering process. Experiment results show that the proposed method is feasible and can yield the better clustering effect.