Abstract

Tree-based boosting models, such as gradient boosting machines (GBMs), have proved to have good predictive performance, at the same time being simple to use for practitioners. Current off-the-shelf GBM program-packages provide natural alternatives to, e.g., GLMs and GAMs. These implementations, and the original GBM-formulation, focus on univariate models, modelling an unknown mean function. In this talk I will discuss how to implement and tune cyclic multidimensional GBMs (CGBMs) that can be used for distributional regression, where you estimate an unknown multidimensional parameter function depending on covariates. That is, CGBMs can be used to jointly model, e.g., the shape and rate parameters in a Gamma-distribution as functions of covariates. A different type of application of the CGBM is to use it to construct more interpretable univariate models: By using the CGBM we introduce an univariate tree-based varying coefficient model, which combines the predictive performance of GBMs with an interpretable GLM-like structure.