A general grid-clustering approach

Hierarchical clustering is an important part of cluster analysis. Based on various theories, numerous hierarchical clustering algorithms have been developed, and new clustering algorithms continue to appear in the literature. It is known that both divisive and agglomerative clustering algorithms in hierarchical clustering play a pivotal role in data-based models, and have been successfully applied in clustering very large datasets. However, hierarchical clustering is parameter-sensitive. When the user has no knowledge of how to choose the input parameters, the clustering results may become undesirable. In this paper, we propose a general grid-clustering approach (GGCA) under a common assumption about hierarchical clustering. The key features of the GGCA include: (1) the combination of the divisible and the agglomerative clustering algorithms into a unifying generative framework; (2) the determination of key input parameters: an optimal grid size for the first time; and (3) the application of a two-phase merging process to aggregate all data objects. Consequently, the GGCA is a non-parametric algorithm which does not require users to input parameters, and exhibits excellent performance in dealing with not well-separated and shape-diverse clusters. Some experimental results comparing the proposed GGCA with the existing methods show the superiority of the GGCA approach.