SVM-based feature selection of latent semantic features

Latent Semantic Indexing (LSI) is an effective method to extract features that captures underlying latent semantic structure in the word usage across documents. However, subspace selected by this method may not be the most appropriate one to classify documents, since it orders extracted features according to their variances, not the classification power. We propose to apply feature ordering method based on support vector machines in order to select LSI-features that is suited for classification. Experimental results suggest that the method improves classification performance with considerably more compact representation. (C) 2004 Elsevier B.V. All rights reserved.