一种初始化不敏感的谱聚类算法

An Initialization Insensitive Spectral Clustering Algorithm

针对传统谱聚类算法初始化敏感引起的聚类效率与正确率不稳定问题,给出一种改进的谱聚类算法.该算法首先构造Laplacian矩阵并得到其特征谱空间,然后引入粒子群优化的FCM算法在该空间中寻找最优粒子作为初始类簇中心用以解决敏感问题.实验表明,与传统谱聚类算法比较,该算法的聚类结果更稳定,在较高维数据集上聚类效率与正确率有明显提高.

Due to the problem that the efficiency and accuracy of original spectral clustering is unstable because of its initialization sensitive,this paper presents an improved spectral clustering algorithm.The improved algorithm first constructs Laplacian matrix and its spectral eigenspace,then introduces the particle swarm optimized（PSO） FCM algorithm to find the optimum swarm and use the swarm as initial cluster centers to solve the sensitive problem.Experiment shows that the clustering result of this improved algorithm is more stable,the efficiency and accuracy of this algorithm in high-dimensional data sets are better than original spectral clustering algorithm.