On the Two-level Hybrid Clustering Algorithm
Eng Yeow Cheu, Chee Keong Kwoh, Zonglin Zhou · 2004
In this paper, we design the hybrid clustering algorithms, which involve two level clustering. At each of the levels, users can select the k-means, hierarchical or SOM clustering techniques. Unlike the existing cluster analysis techniques, the hybrid clustering approach developed here represents the original data set using a smaller set of prototype vectors (cluster means), which allows efficient use of a clustering algorithm to divide the prototype into groups at the first level. Since the clustering at the first level provides data abstraction first, it reduces the number of samples for the second level clustering. The reduction of the number of samples, hence, the reduction of computational cost is especially important when hierarchical clustering is used in the second stage. The prototypes clustered at the first level are local averages of the data and therefore less sensitive to random variations than the original data. The empirical evaluation of the two-level hybrid clustering algorithms is made at four data sets 1.