Case studies in divisive hierarchical clustering

摘要

Hierarchical clustering algorithms can be characterised as agglomerative or divisive. Divisive clustering algorithms are not as popular as agglomerative, due to the complexity involved in their processes. The divisive algorithm Divisive ANAlysis (DIANA) was proposed as an attempt to minimise the computational complexity embedded in divisive algorithms. This work focuses on an empirical comparative analysis of clustering results obtained by four algorithms, DIANA, k-means, EM and farthest first, in nine sets of patterns having different characteristics. Taking into account the sets of patterns used in the experiments, in general, the clusterings induced by the DIANA algorithm are an indication that the bisection strategy employed by the algorithm can produce clustering entirely different from those produced by the other three algorithms. Its splits are performed recursively down the clustering hierarchy, in a greedy way - once it is done, there is no way back to restore a previous cluster.