School of Information Engineering and Automation,Kunming University of Science and Technology
选取合理的初始聚类中心是正确聚类的前提, 针对现有的K-means 算法随机选取聚类中心和无法处理离群点等问题, 提出一种基于相异性度量选取初始聚类中心改进的K-means 聚类算法. 算法根据各数据对象之间的相异性构造相异性矩阵, 定义了均值相异性和总体相异性两种度量准则; 然后据此准则来确定初始聚类中心, 并利用各簇中数据点的中位数代替均值进行后续聚类中心的迭代, 消除离群点对聚类准确率的影响. 此外, 所提出的算法每次运行结果保持一致, 在初始化和处理离群点方面具有较好的鲁棒性; 最后, 在人工合成数据集和UCI数据集上进行实验, 与3 种经典聚类算法和两种优化初始聚类中心改进的K-means 算法相比, 所提出的算法具有较好的聚类性能.
Selecting a reasonable initial clustering center is the premise of correct clustering. Most of the existing K-means algorithms have some shortcomings, such as randomly selecting clustering centers and unable to deal with outliers, an improved K-means clustering algorithm for selecting initial clustering centers based on dissimilarity measure is proposed. According to the dissimilarity of each data objects, the dissimilarity matrix is constructed, and two measures of mean dissimilarity and total dissimilarity are defined. Then the initial clustering center is determined according to the criteria, and the median of data points in each cluster is used to replace the mean value for the subsequent iteration of clustering center, so as to eliminate the effect of outliers on clustering accuracy. In addition, the proposed algorithm maintains consistent results every time, and has better robustness in initializing and handling outliers. Finally, experiments are performed on the synthetic datasets and UCI datasets. Compared with three classical clustering algorithms and two improved K-means algorithms, the proposed algorithm has better clustering performance.