Highly Scalable Discriminative Spam Filtering
Michael Brückner, Peter Haider, Tobias Scheffer · 2008
This paper discusses several lessons learned from the SpamTREC 2006 challenge. We discuss issues related to decoding, preprocessing, and tokenization of email messages. Using the Winnow algorithm with orthogonal sparse bigram features, we construct an efficient, highly scalable incremental classifier, trained to maximize a discriminative optimization criterion. The algorithm easily scales to millions of training messages and millions of features. We address the composition of training corpora and discuss experiments that guide the construction of our SpamTREC entry. We describe our submission for the filtering tasks with periodical re-training and active learning strategies, and report on the evaluation on the publicly available corpora.