A rough sets-based rule induction for numerical datasets

Abstract

Drawing reasonable conclusions from real-world data has been a challenge owing to diverse factors related to the quality of information. In order to handle these problems, the rough sets theory, which deals with inconsistency through the approximation of data sets, was proposed. Among the applications of rough sets, their use in learning processes is highlighted due to their capacity to produce interpretable classification models. Despite their success, some of the most commonly used rough sets based methods are designed to work with categorical input data. This design choice can severely limit their application to real-world problems. Such methods are also inappropriate to handle binary classification problems. By using discretization methods and decision trees we were able to overcome such limitations and improve the classification quality of the methods used. As a result, we developed three approaches. The first presented approach produce interpretable results considering the rejection option. The second approach makes use of belief merging techniques to reduce the number of rejected objects in the first approach. At last, the third approach makes use of decision trees to classify all rejected cases.