An improved selective ensemble method for spam filtering

This paper presents an improved method of selective ensemble to filter the spam messages. The design adopts clustering based on the diversity between sub-classifiers to solve the problem of selection. To improve accuracy and stability, a conception of confidence weight is proposed to evaluate the reliability of selected sub-classifiers. The training model is created with small datasets as in the real situation. For practical usage, this method only uses 150 samples of user's file and executes bootstrapping between 50 and 70 times on them. Experiments validate the effectiveness of this method in handling the spam filtering problem.