Pairwise Multi-Class Feature Selection: A Wrapper Framework
Fred, A. L. N.
WSEAS Trans. on Systems and Control Vol. 2, Nº 2, pp. 212 - 217, February, 2007.
Journal Impact Factor: (in 0)
Digital Object Identifier:
Wrapper feature selection methods are typically used in multi-class classification problems to determine which feature subspace maximizes the patterns discriminative potential, with respect to the global multi class scope. However, in most classification tasks, some classes are more easily discriminated than others due to par- ticularly predictive features. Thus the global class set may stand as a hard restriction when performing feature selection. We propose a class pairwise approach, in which the wrapper feature selection framework is applied with the purpose of determining the feature subspaces with higher discriminative potential for each class pair. This method is shown to provide simpler models, reduced number of features, higher scalability, and in some cases even improve the classification performance.