The design of a recognition system requires careful attention to pattern representation and classifier design. Some statistical approaches choose those features, in a d-dimensional initial space, which allow sample vectors belonging to different categories to occupy compact and disjoint regions in a low-dimensional subspace. The effectiveness of the representation subspace is then determined by how well samples from different classes can be separated. In this paper, we propose a feature selection process that sorts the principal components, generated by principal component analysis, in the order of their importance to solve a specific recognition task. This method provides a low-dimensional representation subspace which has been optimized to improve the classification accuracy. We focus on the problem of facial expression recognition to demonstrate this technique. We also propose a decision tree-based classifier that provides a ''coarse-to-fine'' classification of new samples by successive projections onto more and more precise representation subspaces. Results confirm, first, that the choice of the representation strongly influences the classification results, second that a classifier has to be designed for a specific representation.