Using BP-network to construct fuzzy decision tree with composite attributes

Decision trees have the characteristics of quick learning while its ability of representing complex concept is not very strong. This paper proposes a new algorithm, which can improve the ability of concept representation of decision trees, by utilizing a neural network to compound composite attributes. The composite attribute is compounded only when needed. The computational complexity is not increased much. Experiments show that both training accuracy and test accuracy have remarkable improvements by using this algorithm.