基于LPCA的谱聚类算法

针对传统谱聚类在构建关系矩阵时只考虑样本的全局特征而忽略样本的局部特征、在聚类划分时通常需要指定聚类个数、无法对交叉点进行正确划分等问题，提出了一种改进的基于局部主成分分析和连通图分解的谱聚类算法。首先自动学习挑选数据集的中心点，然后使用局部主成分分析得到数据集的关系矩阵，最后用连通图分解算法完成对关系矩阵的划分。实验结果表明提出的改进算法性能优于现有经典算法。

Spectral clustering algorithm based on LPCA

As the traditional spectral clustering algorithms only considered the global structures of the samples while ignoring their local structures for the construction of the correlation matrix; conducted clustering with a predefined cluster number; could not divide the intersections correctly. This paper proposed a new method based on the local principal component analysis and the decomposition method of the connected graph. Specifically, the proposed method automatically learnt the centroids of the selected subset of the samples, obtained the correlation matrix of the samples based on the local principal component analysis, and used the decomposition method of the connected graph to partition the resulting correlation matrix. Experimental results show that the proposed algorithm performs better than the existing algorithms.