On the Impact of Document Representation on Classifier Per-formance in e-Mail Categorization.
Helmut Berger, Monika Köhle, Dieter Merkl · ISTA · 2005
Abstract: Thispaperprovidesananalysisofmulti-classe-mailcategorizationper-formance.Inordertoinvestigatethisissue,thequalityofvariousclassificational-gorithmsbasedontwodistinctdocumentrepresentationformalismsiscompared.Inparticular,bothastandardword-baseddocumentrepresentationaswellasacharactern -gramdocumentrepresentationisused.Thelatterisregardedashighlynoise-tolerantandwasoriginallyproposedforautomaticlanguageidentification andasaconvenientmeansforproducingcompactdocumentindices. Furthermoretheimpactofusingavailablee-mailspecificmeta-informationonclassificationperformanceisexploredandthefindingsarepresented. 1Introduction Thetaskofautomaticallysortingdocumentsofadocumentcollectionintocategoriesfromapredefinedset,isreferredtoastext categorization .Textcategorizationisapplicableinavarietyofdomains:documentgenreidentification,authorshipattribution,surveycoding,tonamebutafew[Seb02].Oneparticularapplicationiscategorizinge-mailmessagesintolegitimateandspammessages,i.e.spam filtering.Androutsopoulosetal.comparein[APK