Summary
In pattern recognition, many learning methods need numbers as inputs. This paper analyzes two-level classifier ensembles to improve numeric learning methods on nominal data. A different classifier was used at each level. The classifier at the base level transforms the nominal inputs into continuous probabilities that the classifier at the meta level uses as inputs. An experimental validation is provided over 27 nominal datasets for enhancing a method that requires numerical inputs (e.g. Support Vector Machine, SVM). Cascading, Stacking and Grading are used as two-level ensemble implementations. Experiments combine these methods with another symbolic-to-numerical transformation — Value Difference Metric, VDM. The results suggest that Cascading with Binary Decision Trees at base level and SVM with VDM at meta level produces a better accuracy than other possible two-level configurations.