Feature Selection by Means of a Feature Weighting Approach

Selecting a set of features which is optimal for a given classiication task is one of the central problems in machine learning. We address the problem using the exible and robust lter technique EUBAFES. EUBAFES is based on a feature weighting approach which computes binary feature weights and therefore a solution in the feature selection sense and also gives detailed information about feature relevance by continuous weights. Moreover the user gets not only one but several potentially optimal feature subsets which is important for lter-based feature selection algorithms since it gives the exibility to use even complex classiiers by the application of a combined lter/wrapper approach. We applied EUBAFES on a number of artiicial and real world data sets and used radial basis function networks to examine the impact of the feature subsets to classiier accuracy and complexity.