面向大规模数据的分层近邻传播聚类算法

Hierarchical Affinity Propagation Clustering for Large-scale Data Set

近邻传播(Affinity Propagation,AP)聚类具有不需要设定聚类个数、快速准确的优点,但无法适应于大规模数据的应用需求。针对此问题,提出了分层近邻传播聚类算法。首先,将待聚类数据集划分为若干适合AP算法高效执行的子集,分别推举出各个子集的聚类中心;然后对所有子集聚类中心再次执行AP聚类,推举出整个数据集的全局聚类中心;最后根据与这些全局聚类中心的相似度对聚类样本进行划分,从而实现对大规模数据的高效聚类。在真实和模拟数据集上的实验结果均表明,与AP聚类和自适应AP聚类相比,该方法在保证较好聚类效果的同时,极大地降低了聚类的时间消耗。

Affinity Propagation （AP） has advantages on efficiency and accuracy,and has no need to set the number of clusters,but is not suitable for large-scale data clustering.Hierarchical Affinity Propagation （HAP） was proposed to overcome this problem.Firstly,the data set was divided into several subsets that can be effectively clustered by AP to select the exemplars of each subset.Then,AP clustering was implemented again on all the subset exemplars to select exemplars of the whole data set.Finally,all the data points were clustered according to similarities with the exemplars,and realizing efficient clustering of large-scale data set.The experimental results on real and simulated data sets show that,compared with traditional AP and adaptive AP,HAP reduces the time consumption greatly and achieves a good clustering result in the meanwhile.