A Selective Learning Model For Spam Filtering

We investigate how one can use optimization meth-ods to attune spam filters to the specific issues found in the spam filtering area. Among those issues is the need to modelize and manage spammers strategies to delude the filters. To adress this issue, we propose a selective learning scheme designed to maximize learn-ing efficiency. In an offline context, this model uses a simple metaheuristic approach to select a subpart of training data such that the filter induced on that part maximizes its accuracy over the evaluation set. In an online context, we show how one machine learning algorithm can discard incoming messages in order to prevent its knowledge base to be biased by messages which are not good representatives of their class and thus, may lead to a decrease in accuracy if they were to be learned. We show that this approach synergizes well with existing classification models while increas-ing significantly their efficiency over time. More im-portantly, we show that this model make existing fil-ters less vulnerable to spammers attempts to delude them