The research work described in the paper addressed the preparation of the input data for the classical rough set approach, with the aim of preserving the informative content of all condition attributes. Instead of ignoring attributes whose values are assigned to a single interval by a supervised discretisation algorithm, such attributes are subjected to unsupervised discretisation processing. In order to examine the informativeness of attributes undergoing the fusion of discretisation methods, reducts and decision rules were induced as popular forms of knowledge representation, especially in the framework of rough set theory. The results obtained were studied from the point of view of the characteristics of knowledge representations and the performance of rule-based classifiers evaluated with test sets discretised in different ways. The conducted experiments demonstrate the validity of the investigated approach.