Using Pairs of Data-Points to Define Splits for Decision Trees
Geoffrey E. Hinton and Michael Revow
Department of Computer Science
University of Toronto
Conventional binary classification trees such as CART either split
the data using axis-aligned hyperplanes or they perform a computationally expensive search
in the continuous space of hyperplanes with unrestricted orientations. We show that the
limitations of the former can be overcome without resorting to the latter. For every pair
of training data-points, there is one hyperplane that is orthogonal to the line joining
the data-points and bisects this line. Such hyperplanes are plausible candidates for
splits. In a comparison on a suite of 12 datasets we found that this method of generating
candidate splits outperformed the standard methods, particularly when the training sets
were small.
Download [Postscript] [pdf]
Advances in Neural Information Processing Systems 8. D.S.
Touretzky, M.C. Mozer and M.E. Hasselmo. MIT Press.
[home page] [publications]