Adopting Hierarchial Cluster Analysis to Improve The Performance of K-mean Algorithm

Abstract

Due to inadequate flood data it is not always possible to fit a frequency analysis to at-site stations. Reliable results are not always guaranteed by a single clustering algorithm, so a combination of methods may be used. In this research, we considered three clustering algorithms: single linkge, complete linkage and Ward (as hierarchial clustering methods), and K-mean (as partitional clustering analysis). Hybrid cluster analysis was tested for up-to-dated of floods data in 68 hydrometric stations in East and NE of Iran. Four cluster validity indices were used to find the optimum number of clusters. Based on the Cophenetic coefficient and average Silhouette width, single linkge, and complete linkage methods were performed well, yet they produced non-consistent clusters (one large and numerous small clusters) which are not amenable for flood frequency analysis. It was shown that hybridization was efficient to form homogeneous regions, however, the usefulness was dependent to the number of classes. Heterogeneity measure of Hosking was negative, due to inter-correlation of floods in the clusters. The hybrid of Ward and K-mean was shown to be the best combination for the region under study. Four homogeneous regions were delineated.

Keywords

• Cluster analysis
• Hyrid
• Khorasan
• Linear moments
• Regional flood frequency analysis
• Regionalyzation