Rule Induction With Probabilistic Rough Classifiers

We present an algorithm ProbRough for inducing decision rules from data. The algorithm combines all the positive aspects of rule induction systems with the exibility of the probabilistic representation of data and learned knowledge. It accepts imprecise and incomplete data of any mixed qualitative and quantitative type. The probabilistic structure of data is estimated from the learning set using the frequency-based estimators. The results of learning are described in the form of a certain partition of the feature space. The partition elements are of a special geometrical shape, which enables the presentation of the decision rules in a simple form meaningful to humans. Nevertheless, the prediction accuracy is comparable or superior to more sophisticated classiiers. The ProbRough algorithm searches through the set of various partitions using the criterion based on minimizing the misclassiica-tion costs. We illustrate the algorithm behavior on several real-life datasets, acknowledged in the machine learning community.