基于稠密区域的K-medoids聚类算法

Novel K-medoids clustering algorithm based on dense regional block

针对传统K-medoids聚类算法对初始中心点敏感,以及迭代次数较高等缺点,提出一种可行的初始化方法和中心点搜索更新策略。新算法首先利用密度可达思想为数据集中每个对象建立一个稠密区域,遴选出K个密度大且距离较远的稠密区域,把对应的稠密区域的核心对象作为聚类算法的K个初始中心点;其次,把K个中心点搜索更新范围锁定在所选的K个有效稠密区域里。新算法在Iris、Wine、PId标准数据集中测试,获取了理想中心点和稠密区域,并且在较少的迭代次数内收敛到最优解或近似最优解。

In view of the traditional K-medoids clustering algorithm is sensitive to the initial center, as well as the shortcomingof high number of iterations, put forward a feasible initialization method and a center search update strategy. New algorithmfirstly using the density-reachable thought to establish a dense regional block for each object of the data set, select Kdense regional blocks which their densities are larger and the distance are far away for each selected dense regionalblocks, put the core object of the corresponding dense regional blocks as the K initial centers;Secondly, the centers searchupdate scope is locking the K selected effective dense regional blocks. Tested on Iris, Wine and PId standard data sets,this new algorithm obtains ideal initial centers and dense regional blocks, what’s more, converges to the optimal solutionor approximate optimum solution within less number of iterations.