Classification with the CTW algorithm

In this paper, we provide a unified view of generative classification and sequence probability estimation. The context-tree weighting (CTW) algorithm, which was originally proposed as an optimum universal source coding algorithm, is thus ex tended to the field of pattern recognition. While focusing on binary classification problems in this paper, we propose single tree and multiple tree CT’~V classifiers, and derive decision criteria for both classifiers. A few ad-hoc approaches to order the features in a single tree and divide the features into different trees are pre sented. As an example in our numerical study, we consider spain email filtering problem. Numerical results show that the approaches proposed in this paper clearly outperforms the naive Bayes classifier, which is a conventional generative classifier.