Abstract
In empirical data modelling, a model of system is built up from a set of cases that the system has observed. Eventually, the performance of the inducted model is dominated by the quality and quantity of observations. Feature transformation methods are widely used to improve quality of knowledge extracted from observations to build up more accurate and robust model. In the paper, a new feature transformation method named dynamical feature bundling for decision tree algorithm is proposed. Dynamical feature bundling groups a set of features in the tree induction phase and it enables decision tree algorithms to 1) make use of features in one bundle together to make collective judgments in splitting phase; 2) learn more reliable and stable knowledge from feature bundles created based on domain knowledge of experts; 3) embed feature transformation step into tree induction phase, and therefore the extra pre-process step which are necessary for static feature transformation methods is inessential. Our experiments show 2%-9% improvements of AUC value on a very imbalanced dataset. Slight improvements are also obtained on a more balanced data set.