随着聚类分析对象数据集规模的急剧增大,改进已有的算法以获得满意的效率受到越来越多的重视.讨论了一类采用数据空间网格划分的基于密度的聚类算法的均值近似方法.该方法过滤并释放位于稠密超方格中的数据项,并利用其重心点近似计算其对周围数据元素的影响因子.给出均值近似在聚类算法中的实现策略及其误差估计.均值近似方法在有效减少内存需求、大幅度降低计算复杂度的同时对聚类精确度影响甚微.实验结果验证了该方法能够取得令人满意的效果.

In recent years, the explosively growing amount of data in numerous clustering tasks has attracted considerable interest in boosting the existing clustering algorithms to large datasets. In this paper, the mean approximation approach is discussed to improve a spectrum of partition-oriented density-based algorithms. This approach filters out the data objects in the crowded grids and approximates their influence to the rest by their gravity centers. Strategies on implementation issues as well as the error bound of the mean approximation are presented. Mean approximation leads to less memory usage and simplifies computational complexity with minor lose of the clustering accuracy. Results of exhaustive experiments reveal the promising performance of this approach.

(Supported by the National Natural Science Foundation of China under Grant No.79970092 (国家自然科学基金); the Natural Science Foundation of the Education Board of Jiangsu Province of China under Grant No.02KJB520012 (江苏省教育厅自然科学基金)