Optimized Representation for Classifying Qualitative Data

Extracting knowledge out of qualitative data is an ever-growing issue in our networking world. Opposite to the widespread trend consisting of extending general classification methods to zero/one-valued qualitative variables, we explore here another path: we first build a specific representation for these data, respectful of the non-occurrence as well as presence of an item, and making the interactions between variables explicit. Combinatorics considerations in our Midova expansion method limit the proliferation of itemsets when building level k+1 on level k, and limit the maximal level K. We validate our approach on three of the public access datasets of University of California, Irvine, repository: our generalization accuracy is equal or better than the best reported one, to our knowledge, on Breast Cancer and TicTacToe datasets, honorable on Monks-2 near-parity problem.