Mining quantitative class-association rules for software size estimation

Associative models are usually applied in knowledge discovery problems in order to find patterns in large databases containing mainly nominal data. This work is focused on two different aspects, the predictive use of association rules and the management of quantitative attributes. The aim is to induce class association rules that allow predicting software size from attributes obtained in early stages of the project. In this application area, most of the attributes are continuous; therefore, they should be discretized before generating the rules. Discretization is a data mining preprocessing task having a special importance in association rule mining since it has a significant influence on the quality and the predictive precision of the induced rules. In this paper, a multivariate supervised discretization method is proposed, which takes into account the predictive purpose of the association rules.