Spam filtering : optimization approaches to content-based filtering
Didier Colin · OpenGrey (Institut de l'Information Scientifique et Technique) · 2009
Spam filtering is a problem which have drawn the attention of the academic world in the early 2000s. While it is mostly viewed as a supervised classification problem, spam filtering brings issues which are not well addressed by a machine learning approach : adversarial classification, or the need for a filter to include the existence of an aware adversary in its classification process, cost-sensitive classification, and the need to minimize human assistance in the learning process, especially in an online context. The purpose of this thesis is to address these issues by bringing an optimization approach to the spam filtering problem. Viewing classifiers as structures to optimize, we formulate the learning processus as an optimization problem, on which we propose to apply a meta-heuristic method, allowing for the induction of more efficient and autonomous filters. Our work also lead us to explore alternative paradigms for spam filtering (social network analysis, game theoretic models), and their association in a unified filtering system. Finally, we propose the spamtools java package, a library designed to ease the implementation of experimental filters, and their interfacing with standardized evaluation tools such as the TREC evaluation toolkit.