Feature selection in robust clustering based on Laplace mixture

A wrapped feature selection process is proposed in the context of robust clustering based on Laplace mixture models. The clustering approach we consider is a generalization of the K-median algorithm. The selection process makes use of the statistical model and recursively deletes features using hypothesis tests. We report simulations and applications to real data sets which illustrate the relevance of the proposed approach. We propose a strategy to select a reasonable number of remaining features. It uses the test statistic to choose the most relevant features, then an evaluation of the clustering error to discard the redundant ones from among them. This strategy appears to produce a good compromise between the selection of features and the performance of the clustering.