In this paper we present the supervised iterative projections and rotations (S-IPR) algorithm, a method to optimise a set of discriminative subspaces for supervised classification. We show how the proposed technique is based on our previous unsupervised iterative projections and rotations (IPR) algorithm for incoherent dictionary learning, and how projecting the features onto the learned sub-spaces can be employed as a feature transform algorithm in the context of classification. Numerical experiments on the FISHERIRIS and on the USPS datasets, and a comparison with the PCA and LDA methods for feature transform demonstrates the value of the proposed technique and its potential as a tool for machine learning.