Greedy selection of attributes to be discretised

It is well known that discretisation of datasets in some cases may improve the quality of a decision system. Such effects were observed many times during experiments conducted in stylometry domain when authorship attribution tasks were performed. However, some experiments delivered results worse than expected when all attributes in datasets were discretised. Therefore, the idea to test decision systems where only part of attributes is discretised arose. For the selection of attributes to be discretised the greedy forward and backward sequential selection methods were proposed and deeply investigated. Different supervised and unsupervised discretisation methods were employed. The Naive Bayes classifier was selected as the inducer in the decision system. The relation between the subsequent subsets of attributes being discretised and the performance of the decision system was observed. The research proved that there is the maximum of the measure of system quality in respect to the series of subsets of attributes being discretised, generated during the sequential selection processes. Therefore, the attempts to find the optimal subsets of attributes to be discretised are reasonable.