对k-means初始聚类中心的优化

Optimization to k-means initial cluster centers

针对传统k-means算法对初始聚类中心敏感的问题,提出了基于数据样本分布选取初始聚类中心的改进k-means算法。该算法利用贪心思想构建K个数据集合,集合的大小与数据的实际分布密切相关,集合中的数据彼此间相互靠近。取集合中数据的平均值作为初始聚类中心,由此得到的初始聚类中心非常接近迭代聚类算法期待的聚类中心。理论分析和实验结果表明,改进算法能改善其聚类性能,并能得到稳定的聚类结果,取得较高的分类准确率。

To solve this problems that the traditional k-means algorithm has sensitivity to the initial cluster centers, an new improved k- means algorithm is proposed. A new method for selecting initial cluster centers based on the actual distribution of data is presented. First it constructes K data sets using greedy thinking, the sizes of sets are related to the real distribution of the entire data. The data belong to the same set are close to each other. Then taking the average of the data sets as the initial cluster centers. The initial cluster centers are found to be very close to the desired cluster centers for iterative clustering algorithms. Theory analysis and experimental results demon- strate that the improved algorithm can enhance the clustering performance, get stable clustering in a higher accuracy.