Distance Plot ( B ) Size Plot ( C ) Time

Clustering is an important data exploration task. A prominent clustering algorithm is agglomerative hierarchical clustering. Roughly, in each iteration, it merges the closest pair of clusters. It was rst proposed way back in 1951, and since then there have been numerous mod-iications. Some of its good features are: a natural, simple, and non-parametric grouping of similar objects which is capable of nding clusters of diierent shape such as spherical and arbitrary. But large CPU time and high memory requirement limit its use for large data. In this paper we show that geometric metric (centroid, median, and minimum variance) algorithms obey a 90-10 relationship where roughly the rst 90iterations are spent on merging clusters with distance less than 10the maximum merging distance. This characteristic is exploited by partially overlapping partitioning. It is shown with experiments and analyses that diierent types of existing algorithms beneet excellently by drastically reducing CPU time and memory. Other contributions of this paper include comparison study of multi-dimensional visa -vis single-dimensional partitioning, and analytical and experimental discussions on setting of parameters such as number of partitions and dimensions for partitioning.