An Approach to Mining Data with Continuous Decision Values

We propose a novel approach to discover useful patterns from ill-defined decision tables with a real value decision and nominal conditional attributes. The proposed solution is based on a two-layered learning algorithm. In the first layer the preference relation between objects is approximated from the data. In the second layer the approximated preference relation is used to create three applications: (1) to learn a ranking order on a collection of combinations, (2) to predict the real decision value, (3) to optimize the process of searching for the combination with maximal decision.