Multiple data structure discovery through global optimisation, meta clustering and consensus methods

When dealing with real data, clustering becomes a very complex problem, usually admitting many reasonable solutions. Moreover, even if completely different, such solutions can appear almost equivalent from the point of view of classical quality measures such as the distortion value. This implies that blind optimisation techniques alone are prone to discard qualitatively interesting solutions. In this work we propose a systematic approach to clustering, including the generation of a number of good solutions through global optimisation, the analysis of such solutions through meta clustering and the final construction of a small set of solutions through consensus clustering.