Efficient Semantic Kernel-Based Text Classification Using Matching Pursuit KFDA.

Qing Zhang, Jianwu Li

Abstract: A number of powerful kernel-based learning machines, such as support vector machines (SVMs), kernel Fisher discriminant analysis (KFDA), have been proposed with competitive performance. However, directly applying existing attractive kernel approaches to text classification (TC) task will suffer semantic related information deficiency and incur huge computation costs hindering their practical use in numerous large scale and real-time applications with fast testing requirement. To tackle this problem, this paper proposes a novel semantic kernel-based framework for efficient TC which offers a sparse representation of the final optimal prediction function while preserving the semantic related information in kernel approximate subspace. Experiments on 20-Newsgroup dataset demonstrate the proposed method compared with SVM and KNN (K-nearest neighbor) can significantly reduce the computation costs in predicating phase while maintaining considerable classification accuracy.