Document representation for one-class SVM

Previous studies have shown that one-class SVM is a rather weak learning method for text categorization problems. This paper points out that the poor performance observed before is largely due to the fact that the standard term weighting schemes are inadequate for one-class SVMs. We propose several representation modifications, and demonstrate empirically that, with the proposed document representation, the performance of one-class SVM, although trained on only small portion of positive examples, can reach up to 95% of that of two-class SVM trained on the whole labeled dataset.