B1330
Title: Model-based and fuzzy clustering algorithms: A comparative simulation study
Authors: Maria Brigida Ferraro - Sapienza University of Rome (Italy) [presenting]
Marco Alfo - University La Sapienza, Rome (Italy)
Paolo Giordani - Sapienza University of Rome (Italy)
Abstract: Model-based and fuzzy clustering (algorithms) are widely used soft clustering methods. In both cases the objects are not assigned to the clusters via a hard rule allocation. In the former approach, it is assumed that the data are generated by a mixture of probability distributions (usually multivariate Gaussian) in which each component represents a different group or cluster. Each object is ex-post assigned to the clusters using the so-called posterior probability of component membership. In the latter case, no probabilistic assumptions are made and each object belongs to the clusters with a fuzzy membership degree, taking values in $[0,1]$, based on the distances between the objects and the cluster prototypes. Therefore it is quite obvious that the posterior probability of component membership may play a role similar to the membership degree. The aim is at comparing the performance of both approaches by means of a simulation study, also considering robust variants of both clustering approaches. In detail, in the model-based context, also finite mixtures of $t$ distributions with or without trimming are investigated, whilst, in the fuzzy context, the problem of outliers is dealt by means of the so-called underlying noise cluster.
