Combination of ROSVM and LR for Spam Filter

Spam filter benefits from two state-of-the-art discriminative models: Logistic Regression (LR) and Relaxed Online Support Vector Machine (ROSVM). It is natural that two models reach their optimal performance after different training examples. We presented a combination model which integrated LR and ROSVM into a unified one. We divided the training process into two phases. In the first phase, LR was used as filtering model to train and learn, at the same time ROSVM accepted the right result to learn. In the second phase, ROSVM was used as filtering model to train after a point which was found in experiments. Experimental results on the public data sets (TREC06-c, TREC06-p, TREC07-p) showed that the combination of ROSVM and LR spam filter gave the better performance than LR filter and ROSVM filter in immediate feedback.