Fuzzy Clustering inParallel Universes
R. Berthold · 2005
We propose a modified fuzzy c-Means algorithm thatoperates on different feature spaces, so-called parallel universes, simultaneously. Themethodassigns membership val- uesofpatterns todifferent universes, which arethenadopted throughout thetraining. Thisleads tobetter clustering results since patterns notcontributing toclustering inauniverse are (completely orpartially) ignored. Theoutcomeofthealgorithm areclusters distributed overdifferent parallel universes, each modeling a particular, potentially overlapping, subset ofthe data. Onepotential target application oftheproposed method is biological dataanalysis wheredifferent descriptors formolecules areavailable butnoneofthembyitself showsglobal satisfactory prediction results. Inthis paperweshowhowthefuzzy c-Means algorithm canbeextended tooperate inparallel universes and illustrate theusefulness ofthis methodusing results onartificial datasets.