Agglomerative and Divisive Approaches to Unsupervised Learning in Gestalt Clusters

Hierarchical clustering algorithms can be agglomerative or divisive, depending on how partitions are formed. Such algorithms have advantages mainly related to the desired level of granularity the partition should have. The work described in this paper approaches two hierarchical algorithms, one agglomerative (and three of its variants) and the other divisive, focusing on their performance in unsupervised learning tasks related to gestalt clusters. Taking into account that the point sets considered are representative of gestalt clusters, the experiments show that the best results have been obtained when the agglomerative approach was used.