Spam filtering based on online ranking logistic regression

Spam filtering is an important issue in Web information processing. Many machine learning methods are utilized to filter spam. Current researches transform the filtering problem into binary classification, in which the optimization target of the classification model is not consistent with 1-AUC, the usual evaluation measurement. The inconsistence results in the deviation of model optimization, which makes a bad influence on filtering results. In this study, spam filtering was transformed into the ranking model through the optimization oriented to 1-AUC with online ranking logistic regression model then proposed to tackle the deviation of the model's score in the online learning module. TONE (train on or near error), re-sampling and weights update methods were used to promote the learning speed in online adjustment of model's parameters. Experiments on open evaluation datasets show that the developed method is better than the traditional online logistic regression model with statistical significance.