A PAC-Bayesian Margin Bound for Linear Classifiers: Why SVMs work

We investigate how the normalization of vectors influences the result of SVMs. 1 Normalization For the theoretical background, please refer to [1]. 2 Experiments We empirically compare the performances of SVMs with or without normalization, taking text categorization as an example. The dataset is 20-newsgroup, which contains 18828 documents after some unsuitable ones. We conducted 5-fold cross-validation. We evaluate performance using averaged F-measures over 20 categories.

• Type: Conference
• Year: 2000
• Where: NIPS
• Authors: Ralf Herbrich, Thore Graepel
• Keywords: NIPS 2000, NIPS 2007, Text Categorization, Theoretical Background, Unsuitable Ones