Validation of an Alternative Neural Decision Tree

A modified version of Neural Decision Tree (NDT) which was initially proposed by Biau is presented in this paper with significantly extended statistical consistency theoretical results. Numerical experiments are performed to investigate several key features of this alternative NDT. Our experimental results demonstrate the attractiveness of NDT over the traditional artificial neural networks by adopting a strategically selected network structure with appropriate initial weights. Moreover, our results indicate that in all cases we have tested, NDT provides higher classification accuracy than a traditional neural network with a comparable degree of freedom. Empirical evidence also suggests that NDT is more stable and robust with respect to hyper-parameters commonly involved in the training of neural networks.