Abstract: This article describes a new non-parametric regression method that extends additive regression techniques to allow modeling of interactions among predictor variables. The proposed models consist of sums of smooth functions of one or more predictor variables. Each term involving more than one predictor is assumed to be a composition of bivariate functions of simpler terms in the model. The method is demonstrated on simulated and real data sets and predictions are compared to those from additive regression models and Friedman's (1991) multivariate adaptive regression spline (MARS) models.
Key words and phrases: Non-parametric regression, adaptive methods, smoothing.