Granular Computing based Comparison of Agglomerative Clustering

Empirical comparison of clustering methods is challenging because it is difficult to set up evaluation indices like supervised learning. In this paper, we propose weaker evaluation indices when at least one target group is given by domain experts. If so, we can introduce rough set based approximation for empirical comparison of dissimilarities and clustering strategies as follows. When a set of target (target set) is given, a level of clustering tree where one branch includes all the targets can be traced with the number of elements included. The pair (#clusters_of_a_level, #elements_of_a_cluster) can be viewed as indices-pair for a given clustering tree. Then, the algorithm for comparison of metric uses twofold comparison: first, it compares the number of partition, and then compares the ratio. Six distances and seven clustering methods were compared by using the above pair. The target items are based on nursing cares necessary for surgical operation of cataracts. Empirical results show that Euclidean distance and the Ward method is the best for obtaining a suitable clinical pathway for all clustering methods.