自动属性加权的K-调和均值聚类算法

K-HARMONIC MEANS CLUSTERING BASED ON AUTOMATED FEATURE WEIGHTING

针对K-调和均值算法中距离度量将所有属性视为相等重要而存在的不足,提出一种利用自动属性加权的改进聚类算法。在算法的目标函数中,用加权欧氏距离替代传统的欧氏距离,并证明了使得算法能够收敛的属性权重更新机制。为进一步提高聚类性能,将粒子群算法融入到改进的属性加权聚类算法中以抑制其陷于局部最优,其中采用聚类中心和属性权重的值同时表示粒子的位置进行寻优。在UCI数据集的测试结果表明,该算法的聚类指标平均提高了约9个百分点,具有更高的聚类准确性和稳定性。

K-harmonic means algorithm has the disadvantage of viewing all features as the same importance in its distance metric. In light of this,we proposed an improved clustering algorithm which takes the advantage of automated feature weighting. In objective function of the algorithm,we replaced the conventional Euclidian distance with the weighted Euclidian distance,and proved the feature weight update mechanism which enables the algorithm to be converged. In order to further improve the clustering performance,we integrated the particle swarm optimisation into the feature weighting clustering algorithm so as to suppress its problem of being trapped into local optimum,in which we used both the centres of clusters and the value of feature weight to represent the position of each particle for optimisation. Tests result on UCI datasets showed that the clustering index of the proposed algorithm has raised about 9 percents,so our method is more accurate and stable.