采用仿射传播的聚类集成算法

Cluster Ensemble Algorithm Using Affinity Propagation

针对K均值聚类随机初始聚类中心导致的聚类结果不稳定问题,提出一种基于仿射传播的聚类集成算法.该算法把每个聚类集成的成员个体结果看成是原始数据的一个属性,然后在其基础上对聚类成员个体的聚类结果进行加权集成,集成算法采用简单高效的仿射传播聚类,并且提出了直接集成、利用平均规范化互信息(NMI)和聚类有效性Silhouette指标进行加权集成.最后,运用Hungarian算法对仿射传播聚类集成的结果进行类别标签的统一和匹配.在加州大学尔湾分校数据集上进行了实验,结果表明,与集成前的K均值聚类及其他聚类集成算法相比,该算法能有效地提高聚类结果的准确性、鲁棒性和稳定性,建立起来的聚类集成算法具有良好的扩展性和灵活性,而且简单有效.

The result of K-means cluster is instable for random initial clustering centers. A cluster ensemble algorithm based on affinity propagation is proposed,where the result of each cluster individual is regarded as a property of the original data. Following the new properties sets, the results of each cluster individual are carried out to a weighted ensemble, and simple and efficient affinity propagation cluster is chosen in the ensemble algorithm. Furthermore the direct ensemble, the ensemble to weighted ensemble from average normalized mutual information （NMI） and cluster validation indexes Silhouette are uniformly proposed. Finally, Hungarian algorithm is employed to unify and match the category labels for the results of affinity propagation cluster. The results of experiments on University of California Irvine data sets show the higher efficiency for improving the accuracy, robustness and stability of cluster results than the K means clustering before combination and the other clustering ensemble algorithms. The clustering ensemble algorithm gets more extendable and flexible.