We describe a new way to deal with feature selection when boosting is used to assess the relevancy of feature subsets. In the context of wrapper models, the accuracy is here replaced as a performance function by a particular exponential criterion, usually optimized in boosting algorithms. A first experimental study brings to the fore the relevance of our approach. However, this new "boosted" strategy needs the construction at each step of many learners, leading to high computational costs. We focus then, in a second part, on how to speed-up boosting convergence to reduce this complexity. We propose a new update of the instance distribution, which is the core of a boosting algorithm. We exploit these results to implement a new forward selection algorithm which converges much faster using overbiased distributions over learning instances. Speed-up is achieved by reducing the number of weak hypothesis when many identical observations are shared by different classes. A second experimental study on the UCI repository shows significantly speeding improvements with our new update without altering the feature subset selection.
展开▼