自适应局部半径的DBSCAN聚类算法

Self-adaptive Local eps DBSCAN

经典的基于密度的聚类方法 DBSCAN算法需要指定邻域半径和最小数据点阈值两个基本参数.这两个参数的确定对聚类结果的影响非常大.目前缺少有效的参数选择确定方法,同时DBSCAN算法在聚类过程中,使用统一的邻域半径参数,使得密度不均匀集上的聚类质量不高.本文提出一种自适应选择局部半径的密度聚类算法(SALE-DBSCAN),通过确定密度峰值点,自适应选择聚类的局部邻域半径,简化了参数选择的过程;通过使用自适应选择的局部邻域半径扩张密度峰值点的邻域进行聚类,提高了聚类结果质量.实验结果表明,本SALE-DBSCAN算法相较其他密度聚类算法的聚类结果更加准确.

DBSCAN requires users set two basic parameters artificially, eps and minPts. They have an important influence for clustering results. At present,there is a lack of effective methods about selecting parameters. Meanwhile ,DBSCAN uses global parameters during clustering, which leads to clustering quality is poor in data sets with different densities. This paper proposed a Self-Adaptive Local eps DBSCAN （SALE-DBSCAN） ,simplifying process of parameters selection by confirming points with density peak and self-adaptive selecting local neighborhood eps of clustering; Improving clustering quality by using self-adaptive local eps expands neighborhoods of points with density peak. The experiment shows our algorithm＇s clustering quality is better than other based-density clustering algorithms.