A probabilistic classification method based on conditional independences

Abstract. We introduce a probabilistic classification method which exploits the conditional independences between the features (random variables). This idea is at the heart of the selection of a special type of informative features. The so called \(t\)-informative features are selected from the great amount of observed features in order to diminish the uncertainty of the classification variable and in the same time avoid the overfitting of the model. We construct a probabilistic function of the selected \(t\)-informative variables which classifies a new entity into one of the classes. We present the efficiency of our algorithm in real-life applications.