Standard approaches to tackle high-dimensional supervised classification often include variable selection and dimension reduction. The proposed methodology combines clustering of variables and feature selection. Hierarchical clustering of variables allows to built groups of correlated variables and summarizes each group by a synthetic variable. Originality is that groups of variables are unknown a priori. Moreover clustering approach deals with both numerical and categorical variables. Among all the possible partitions, the most relevant synthetic variables are selected with a procedure using random forests. Numerical performances are illustrated on simulated and real datasets. Selection of groups of variables provides easier interpretation of results.