A great amount of machine learning techniques have been applied to problems where data is collected over an extended period of time. However, the disadvantage with many real-world applications is that the distribution underlying the data is likely to change over time. In these situations, a problem that many global eager learners face is their inability to adapt to local concept drift. Concept drift in spam is particularly difficult as the spammers actively change the nature of their messages to elude spam filters. Algorithms that track concept drift must be able to identify a change in the target concept (spam or legitimate e-mails) without direct knowledge of the underlying shift in distribution. In this paper we show how a previously successful instance-based reasoning e-mail filtering model can be improved in order to better track concept drift in spam domain. Our proposal is based on the definition of two complementary techniques able to select both terms and e-mails representative of the current situation. The enhanced system is evaluated against other well-known successful lazy learning approaches in two scenarios, all within a cost-sensitive framework. The results obtained from the experiments carried out are very promising and back up the idea that instance-based reasoning systems can offer a number of advantages tackling concept drift in dynamic problems, as in the case of the anti-spam filtering domain.

In this paper we analyse the strengths and weaknesses of the mainly used feature selection methods in text categorization when they are applied to the spam problem domain. Several experiments with different feature selection methods and content-based filtering techniques are carried out and discussed. Information Gain, χ ²-text, Mutual Information and Document Frequency feature selection methods have been analysed in conjunction with Naïve Bayes, boosting trees, Support Vector Machines and ECUE models in different scenarios. From the experiments carried out the underlying ideas behind feature selection methods are identified and applied for improving the feature selection process of SpamHunting, a novel anti-spam filtering software able to accurate classify suspicious e-mails.

Junk e-mail detection and filtering can be considered a cost-sensitive classification problem. Nevertheless, preprocessing methods and noise reduction strategies used to enhance the computational efficiency in text classification cannot be so efficient in e-mail filtering. This fact is demonstrated here where a comparative study of the use of stopword removal, stemming and different tokenising schemes is presented. The final goal is to preprocess the training e-mail corpora of several content-based techniques for spam filtering (machine approaches and case-based systems). Soundness conclusions are extracted from the experiments carried out where different scenarios are taken into consideration.