Dimensionality Reduction Through Unsupervised Features Selection

As the storage technologies evolve, the amount of available data explodes in both dimensions: samples number and input space dimension. Therefore, one needs dimension reduction techniques to explore and to analyse his huge data sets. Many features selection approaches have been proposed for the supervised learning context, but only few techniques are available to address this issue in the unsupervised learning context. Actually, the problem of unsupervised feature selection becomes more difficult as the samples points’ labels disappear. Thus, most of the methods proposed rely on feature correlations and only pairs of variables are considered. In this paper, we extend the wkmeans algorithm proposed by Huang to the self-organizing maps (SOM) framework and we propose a feature selection approach which relies on the weighting coefficients learned during the optimization process. This SOM-based approach addresses the difficult issue of unsupervised feature selection and is ready to handle high dimensional data sets.