DMC.pdf

The TREC 2005 Spam Evaluation Track showed that non-traditional techniques like Prediction by Partial Matching (PPM) performed exceptionally well, while machine learning techniques like Support Vector Machines (SVM) were not represented. To reconcile these results, the authors tested five classification methods on the TREC public corpus using both batch and online evaluation methodologies, and also conducted cross-validation experiments on the Ling-Spam and PU1 datasets. The results revealed substantial differences due to test assumptions, such as batch vs. online training and testing, message ordering, and tokenization. However, methods that performed well at TREC also performed well using established test methods and corpora. Two previously untested methods, Dynamic Markov Compression and logistic regression, compared favorably with competing approaches. The study aimed to investigate the differences between batch and online models, test methods, and the transferability of filters and results between them.