Оценка устойчивости иерархической кластеризации на основе кофенетической корреляции

Abstract

If there are any outliers in the data, hierarchical clustering can produce poor results. In addition, the dendrogram is sensitive to a set of characteristics by which objects are compared. Therefore, it is required to investigate two types of robustness of hierarchical clustering - to a set of objects and to a set of characteristics. For this, an original approach based on the use of bootstrapping is proposed. As an internal criterion for the reliability of hierarchical clustering, the cophenetic correlation coefficient is used. In the simulation study, various methods of hierarchical clustering are compared for robustness of two types. Recommendations are given on the applicability of methods of hierarchical clustering.