Enriched Discretisation: Information Fusion from Supervised and Unsupervised Processing

When data is incomplete and inconsistent, an approximation of concepts can be obtained by applying the rough set theory. The classical approach allows to recognise only nominal attributes, and only nominal classification is possible. To ensure that the inferred rules are of the highest quality, it is beneficial to have access to all available information. The completeness of information can be compromised when in the data preparation stage, discretisation is included as a necessary step. Even when it is performed taking into account the class labels of instances, discretisation can lead to some information loss. The paper illustrates a research framework that extends the transformation of continuous input features into categorical ones. Processing aims to improve the performance of rule-based classifiers created with rough set data mining. The experiments were conducted in the stylometry domain, with the main task of attribution of the authorship. The results obtained suggest that combining supervised discretisation with elements of unsupervised transformations can lead to improved predictions, thus demonstrating the advantages of the proposed research methodology.