A study on parallel versus sequential relational fuzzy clustering methods

Dissertação para obtenção do Grau de Mestre em Engenharia Informática

Relational Fuzzy Clustering is a recent growing area of study. New algorithms have been developed,as FastMap Fuzzy c-Means (FMFCM) and the Fuzzy Additive Spectral Clustering Method(FADDIS), for which it had been obtained interesting experimental results in the corresponding founding works. Since these algorithms are new in the context of the Fuzzy Relational clustering community, not many experimental studies are available. This thesis comes in response to the need of further investigation on these algorithms, concerning a comparative experimental study from the two families of algorithms: the parallel and the sequential versions. These two families of algorithms differ in the way they cluster data. Parallel versions extract clusters simultaneously from data and need the number of clusters as an input parameter of the algorithms, while the sequential versions extract clusters one-by-one until a stop condition is verified, being the number of clusters a natural output of the algorithm. The algorithms are studied in their effectiveness on retrieving good cluster structures by analysing the quality of the partitions as well as the determination of the number of clusters by applying several validation measures. An extensive simulation study has been conducted over two data generators specifically constructed for the algorithms under study, in particular to study their robustness for data with noise. Results with benchmark real data are also discussed. Particular attention is made on the most adequate pre-processing on relational data, in particular on the pseudo-inverse Laplacian transformation.