Topic sequence kernel

Abstract

This paper addresses the problem of classifying documents using the kernel approaches based on topic sequences. Previously, the string kernel uses the ordered subsequence of characters as features and the word sequence kernel is proposed to use words as the subsequences. However, they both face the problem of computational complexity because of the large amount of symbols (characters or words). This paper, therefore, proposes to use sequences of topics rather than characters or words to reduce the number of symbols, thus increasing the computational efficiency. Documents that exhibit similar posterior topic proportions are expected to have similar topic sequence and then should be classified into the same category. Experiments conducted on the Reuters-21578 datasets have proven this hypothesis. © Springer-Verlag 2012.