基于空间结构的符号数据仿射传播算法*

由于符号型数据缺乏清晰的空间结构,很难构造一种合理的相似性度量,从而使诸多数值型聚类算法难以推广至符号型数据聚类.基于此种情况,文中引入一种空间结构表示方法,把符号型数据转化为数值型数据,能够在保持原符号型数据的结构特征的基础上重新构造样本之间的相似度.基于此方法,将仿射传播(AP)聚类算法迁移至符号数据聚类中,提出基于空间结构的符号数据AP算法(SBAP).在UCI数据集中若干符号型数据集上的实验表明,SBAP可以使AP算法有效处理符号型数据聚类问题,并且可以提升算法性能.

Space Structure Based Affinity Propagation Algorithm for Categorical Data

Constructing a reasonable similarity measure is difficult due to the lack of clear space structure in categorical data. Therefore, numerical clustering algorithms can hardly be extended to categorical data clustering. In this paper, a representation method for transforming the categorical data into numerical data is introduced. The similarity between samples is reconstructured and the structure feature of the original categorical data is maintained in the reconstruction process. Based on the data representation method, the affinity propagation(AP) clustering algorithm is migrated to the categorical data clustering. A space structure based AP algorithm for categorical data(SBAP) is proposed. Experimental results on several categorical datasets from the UCI dataset show that the proposed method makes AP algorithm deal with the categorical data clustering problem effectively with a significant improvement in performance.