Abstract:
|
The estimation of statistical parameters is often associated with error estimates as measurements of uncertainty. In the interest of clustering objects on the basis of parameter vectors obtained by fitting a statistical model, the corresponding parameter errors should also be considered in the clustering process. By ignoring the various levels of uncertainty, we inevitably assign equal weights to all parameter estimates. This might result in obtaining imprecise clustering. The majority of previous studies in this context have considered the uncertainty implicitly in clustering. In this work, we present a new dissimilarity measure that explicitly incorporates the uncertainty of parameter estimates by considering the geometrical overlap of confidence regions of the joint parameter estimates. Hence, the assignments of objects to clusters are based on all possible values of the parameters rather than producing results that are pertained to single values obtained from the point estimates. We present the superior performance of the proposed dissimilarity measure over conventional distance measures through extensive simulation studies and a benchmark data set.
|