clustvarsel: Variable Selection for Model-Based Clustering

The selection method uses either a greedy search or headlong search. The greedy search at each step either checks all variables not currently included in the set of clustering variables singly for inclusion into the set or checks all variables in the set of clustering variables singly for exclusion.The headlong search only checks until a variable is included or excluded (i.e. does not necessarily check all possible variables for inclusion/exclusion at each step) and any variable with evidence of clustering below a certain level at any stage is removed from consideration for the remainder of the algorithm. Each variable's evidence for being useful to the clustering given the currently selected clustering variables is given by the difference between the BIC for the model with clustering (allowed to vary over 2 to a maximum number of groups and any of the different covariance parameterizations allowed in mclust) using the set of clustering variables including the variable being checked and the sum of BICs for the model with clustering (allowed to vary over 2 to a maximum number of groups and any of the different covariance parameterizations allowed in mclust) using the set of clustering variables without the variable being checked and the model for the variable being checked being conditionally independent of the clustering given the other clustering variables (this is modeled as a regression of the variable being checked on the other clustering variables).

Downloads: