Optimal partitioning for classification and regression trees

An iterative algorithm that finds a locally optimal partition for an arbitrary loss function, in time linear in N for each iteration is presented. The algorithm is a K-means-like clustering algorithm that uses as its distance measure a generalization of Kullback's information divergence. Moreover, it is proven that the globally optimal partition must satisfy a nearest neighbour condition using divergence as the distance measure. These results generalize similar results of L. Breiman et al. (1984) to an arbitrary number of classes or regression variables and to an arbitrary number of bills. Experimental results on a text-to-speech example are provided and additional applications of the algorithm, including the design of variable combinations, surrogate splits, composite nodes, and decision graphs, are suggested.<
>