Sélection de Caractéristiques pour le Filtrage de Spams

Feature selection is an important step in classification systems. It aims at reducing the number of features while trying to preserve or improve classifier performance. In this article, we propose a machine learning based feature selection approach, in the context of spam filtering which is considered as a text categorization task. Our approach consists in an individual evaluation of each attribute describing the textual messages in order to sort the features and then select a subset according to a performance evaluation that uses bayesian or SVM classifiers (Naive Bayes or Support Vector Machines respectively). We carried out an experimental comparison by testing several combinations which correspond to variations of classifier types, selection strategies (forward/backward) and individual feature evaluation methods, and we obtained interesting results. Indeed, in some cases, we achieved a significant reduction of the feature number in addition to a performance improvement. MOTS-CLÉS : Sélection de caractéristiques, Filtrage de spams, Classification, Apprentissage automatique.