Feature maximization (F-max) is an unbiased quality estimation metric of unsupervised classification (clustering) that favours clusters with a maximal feature F-measure value. In this article we show that an adaptation of this metric within the framework of supervised classification allows efficient feature selection and feature contrasting to be performed. We experiment the method on different types of textual data. In this context, we demonstrate that this technique significantly improves the performance of classification methods as compared with the use of state-of-the art feature selection techniques, notably in the case of the classification of unbalanced, highly multidimensional and noisy textual data gathered in similar classes.