The authors discuss the virtues and flaws of artificial neural networks (ANNs) and univariate decision trees (UDTs) as universal function approximators. They propose the algorithm C-Net for generating multivariate decision trees (MDTs) from ANNs.
There are three stages to the algorithm. First, a single hidden-layered ANN is trained on a suitable training set. Then, the training set is presented again to the ANN, but the outputs of the hidden units become the input feature vector to Quinlan’s C5. The resulting UDT is then converted to an MDT in the original feature space.
The MDT obtained as a result of combining these two technologies is more expressive than the ANN, and more accurate and compact than the UDT generated by C5 alone. Test results on both artificial and real-life data sets are given.
Finally, the authors introduce the concept of recurrent decision trees, and use C-Net to generate them from recurrent neural networks.