Piecewise hierarchical clustering (Q1177260)

The author tackles an interesting problem in hierarchical clustering: how to classify hierarchically a large set of items by partitioning it in small subsets and then merging the resulting small dendrograms in a big one. Reasons for this approach can be the largeness of the data set, exceeding the computer capacities, or the delay with which all items of the set can be made available so that one can start with the first calculations before the last item is available. The paper proposes algorithms to solve this problem in a consistent way but shows that in general this solution is not unique (uniqueness can only be reached under rather stringent conditions). This interesting paper lacks a final conclusion in which a balance would have been made between the time saved by the proposed approach versus the lost accuracy and uniqueness of the obtained dendrogram. Reported results on actual examples could make this point more palpable.