【摘要】 聚类通常被认为是一种无监督的数据分析方法,然而在实际问题中可以很容易地获得有限的样本先验信息,如样本的成对限制信息.大量研究表明,在聚类搜索过程中充分利用先验信息会显著提高聚类算法的性能.首先分析了在聚类过程中仅利用成对限制信息存在的不足,尝试探索数据集本身固有的先验信息——空间一致性先验信息,并提出利用这类先验信息的具体方法.接着,将两类先验信息同时引入经典的谱聚类算法中,提出一种密度敏感的半监督谱聚类算法(density-sensitivesemi-supervisedspectralclusteringalgorithm,简称DS-SSC).两类先验信息在指导聚类搜索的过程中能够起到相辅相成的作用,这使得DS-SSC算法相对于仅利用成对限制信息的聚类算法在聚类性能上有了显著的提高.在UCI基准数据集、USPS手写体数字集以及TREC的文本数据集上的实验结果验证了这一点.更多还原

【Abstract】 Clustering has been traditionally viewed as an unsupervised method for data analysis. In real world application,however,some background prior knowledge can be easily obtained,such as pairwise constraints. It has been demonstrated that constraints can improve clustering performance. In this paper,the drawback of only incorporating pairwise constraints in clustering is firstly analyzed,and then an inherent prior knowledge in data sets,namely space consistency prior knowledge is exploited. The method of utilizing space consistency prior knowledge is also given. Incorporating the two types of prior knowledge into original spectral clustering,a density-sensitive semi-supervised spectral clustering algorithm (DS-SSC) is proposed. Experimental results on UCI (University of California Irvine) benchmark data,USPS (United States Postal Service) handwritten digits and text data from TREC (Text REtrieval Conference) show that the two types of prior knowledge can supplement each other in clustering process,leading to substantial performance enhancement of DS-SSC over other semi-supervised clustering methods which only incorporate pairwise constraints.更多还原