A re-examination of text categorization methods

This paper reports a controlled study with statistical signicance tests on ve text categorization methods: the Support Vector Machines (SVM), a k-Nearest Neighbor (kNN) classi er, a neural network (NNet) approach, the Linear Leastsquares Fit (LLSF) mapping and a Naive B a yes (NB) classier. We focus on the robustness of these methods in dealing with a skewed category distribution, and their performance as function of the training-set category frequency. Our results show that SVM, kNN and LLSF signi cantly outperform NNet and NB when the number of positive training instances per category are small (less than ten), and that all the methods perform comparably when the categories are su ciently common (over 300 instances).