Real-world　applications　often　require　the　classification　of　web　documents　under　the　situation　of　noisy　data.　Support　vector　machines　(SVM)　work　well　for　classification　applications　because　of　their　high　generalization　ability.　But　they　are　very　sensitive　to　noisy　training　data,　which　can　degrade　their　classification　accuracy.　This　paper　presents　a　new　algorithm　to　deal　with　noisy　training　data,　which　combines　support　vector　machines　and　K-nearest　neighbor　(KNN)　method.　Given　a　training　set,　it　employs　K-nearest　neighbor　method　to　remove　noisy　training　examples.　Then　the　remained　examples　are　selected　to　train　SVM　classifiers　for　web　categorization.　Empirical　results　show　that　this　new　algorithm　has　strong　tolerance　of　noise,　and　it　can　greatly　reduce　the　influence　of　noisy　data　on　the　SVM　classifier.　©　2005　IEEE.