Nonparametric Bayesian Classification

Abstract

A Bayesian approach to the classification problem is proposed in which random partitions play a central role. It is argued that the partitioning approach has the capacity to take advantage of a variety of large-scale spatial structures, if they are present in the unknown regression function $f_0$. An idealized one-dimensional problem is considered in detail. The proposed nonparametric prior uses random split points to partition the unit interval into a random number of pieces. This prior is found to provide a consistent estimate of the regression function in the $\L^p$ topology, for any $1 \leq p < \infty$, and for arbitrary measurable $f_0:[0,1] \to [0,1]$. A Markov chain Monte Carlo (MCMC) implementation is outlined and analyzed. Simulation experiments are conducted to show that the proposed estimate compares favorably with a variety of conventional estimators. A striking resemblance between the posterior mean estimate and the bagged CART estimate is noted and discussed. For higher dimensions, a generalized prior is introduced which employs a random Voronoi partition of the covariate-space. The resulting estimate displays promise on a two-dimensional problem, and extends with a minimum of additional computational effort to arbitrary metric spaces.