OPTIMAL DISCRETIZATION AND SELECTION OF FEATURES BY ASSOCIATION RATES OF JOINT DISTRIBUTIONS

In this paper we propose a new method to measure the contribution of discretized features for supervised learning and discuss its applications to biological data analysis. We restrict the description and the experiments to the most representative case of discretization in two intervals and of samples belonging to two classes. In order to test the validity of the method, we measured the abundance of different explanatory models that can be derived from a given set of binary features. We compare the performances of our algorithm with those of popular feature selection methods, over three different publicly available gene expression data sets. The results of the comparison are in favour of the proposed method.