Some Remarks on Feature Ranking Based Wrappers
Mariusz Kubus · University of Lodz Repository (University of Łódź) · 2013
One of the approaches to feature selection in discrimination or regression is learning models using various feature subsets and evaluating these subsets, basing on model quality criterion (so called wrappers). Heuristic or stochastic search techniques are applied for the choice of feature subsets. The most popular example is stepwise regression which applies hillclimbing. Alternative approach is that features are ranked according to some criterion and then nested models are learned and evaluated. The sophisticated tools of obtaining a feature rankings are tree based ensembles. In this paper we propose the competitive ranking which results in slightly lower classification error. In the empirical study metric and binary noisy variables will be considered. The comparison with a popular stepwise regression also will be given.