Visual analytics for the clustering capability of data

• Published in: Science China. Information sciences 2013-05, Vol.56 (5), p.126-139
• Main Authors: Lu, ZhiMao, Liu, Chen, Zhang, Qi, Zhang, ChunXiang, Fan, DongMei, Yang, Peng
• Format: Article
• Language: English
• Subjects: Boundaries; Boundary layer; Cluster analysis; Clustering; Clusters; Computer Science; Data analysis; Data points; Datasets; Fuzzy sets; Information Systems and Communication Service; Outliers (statistics); Research Paper; Strategy; Visualization; 分析数据; 多维数据集; 聚类分析方法; 聚类结果; 能力; 视觉; 集群技术; 高维数据
• ISSN: 1674-733X; 1869-1919
• DOI: 10.1007/s11432-013-4832-7

Clustering analysis is an unsupervised method to find hidden structures in datasets and has been widely used in various fields. However, it is always difficult for users to understand, evaluate, and explain the clustering results in the spaces with dimension greater than three. Although high-dimensional visualization of clustering technology can express clustering results well, it still has significant limitations. In this paper, a visualization cluster analysis method based on the minimum distance spectrum （MinDS） is proposed, aimed at reducing the problems of clustering multidimensional datasets. First, the concept of MinDS is defined based on the distance between high-dimensional data. MinDS can map any dataset from high-dimensional space to a lower dimension to determine whether the data set is separable. Next, a clustering method which can automatically determine the number of categories is designed based on MinDS. This method is not only able to cluster a dataset with clear boundaries, but can also cluster the dataset with fuzzy boundaries through the edge corrosion strategy based on the energy of each data point. In addition, strategies for removing noise and identifying outliers are designed to clean datasets according to the characteristics of MinDS. The experimental results presented validate the feasibility and effectiveness of the proposed schemes and show that the proposed approach is simple, stable, and efficient, and can achieve multidimensional visualization cluster analysis of complex datasets.