Abstract: Partition clustering is sensitive to initialisation. Moreover, it is not always clear whether a unique best solution exists for a given practical problem. A methodology is presented for mapping the space of partition clusters, from which to sample reproducible cluster solutions with good separation measured by second-order statistics and with stability to repeated initialisations. The cluster solutions are reproducible in the sense that multiple independent re-samplings of the initial conditions will return substantially identical partitions, for the same data and clustering algorithm. An extension of the methodology is proposed to reproducibly define a hierarchical tree of cluster solutions for increasing cluster number. The method is evaluated with synthetic data and applied to a previously studied protein expression dataset.
