Improving textual data classification and discrimination using an ad-hoc metric: Application to a famous text discrimination challenge

Labelling maximization (F-max) is an unbiased metric for estimation of the quality of non-supervised classification (clustering) that promotes the clusters with a maximum value of feature F-measure. In this paper, we show that an adaptation of this metric within the supervised classification allows to perform a selection of features and to calculate for each of them a function of contrast. The method is tested on the famous, difficult deemed and ill-balanced Mitterrand-Chirac talk's dataset of DEFT 2005 challenge. We show that it produces extremely important classification performance improvements on this dataset while allowing to clearly isolate the discriminating characteristics of the different classes (i.e. Chirac and Mitterrand profiles).