Fusion of feature selection methods for pairwise scoring SVM

It has been recently discovered that stacking the pairwise comparison scores between unknown patterns and a set of known patterns can result in feature vectors with desirable discriminative properties for classification. However, such technique can be hampered by the curse of dimensionality because the vectors size is equal to the training set size. To overcome this problem, this paper investigates various filter and wrapper feature selection techniques for reducing the feature dimension of pairwise scoring matrices and argues that these two types of selection techniques are complementary to each other. Two fusion strategies are then proposed to (1) combine the ranking criteria of filter and wrapper methods at algorithmic level and (2) merge the features selected by the filter and wrapper methods. Evaluations on a subcellular localization benchmark and a microarray dataset demonstrate that feature subsets selected by the fusion methods are either superior to or at least as good as those selected by the individual methods alone for a wide range of feature dimensions.