空间聚类有效性评价方法对比与研究

A Comparative Study of Validity Evaluation Approaches of Spatial Clustering

空间聚类是当前地球信息科学与计算机科学领域共同关注的热点问题之一,常用来揭示空间数据分布规律以及发现空间数据异常。空间聚类有效性评价即对空间聚类结果进行定量、客观的评判,对于在实际应用中针对不同数据集选取最优的空间聚类算法以及确定最佳的聚类参数具有重要意义。首先选取并编程实现了数种空间聚类有效性评价方法,包括聚类中心的距离矩阵、距离方差、改进Hubert's统计、Davies-Bouldin、Calinski-Harabasz和基于信息论的空间聚类有效性评价方法等,同时提出了顾及簇间分离度和簇内紧凑度的空间聚类有效性评价方法,并通过试验分析验证了其可行性及有效性。然后在K-Means法对数据集进行聚类的基础上,对比研究了前述聚类有效性评价方法的特性及优缺点。

Spatial clustering has been a hot research topic for geographical information science and computer science.It has been used to explore the spatial distribution pattern and identify spatial outliers.The validity evaluation on spatial clustering is about using certain quantitative and objective measurements to evaluate the spatial clustering results.Validity evaluation on spatial clustering is necessary especially for the selection of the optimal spatial clustering algorithm and to define clustering parameters.In this article we select and program several evaluation approaches for spatial clustering including distance matrix of the clustering center,distance variance,improved Hubert'sstatistics,Davies-Bouldin,Calinski-Harabasz,information-based approach,etc.We also propose a new evaluation approach for spatial clustering,which considers disparity between clusters and compactness of a cluster.The newly proposed approach have been evaluated.Then we use K-means approach to obtain spatial clustering results of given point sets.Finally a comparative study of validity evaluation approaches of spatial clustering is carried out.