Decision tree boosted varying coefficient models

Abstract Varying coefficient models are a flexible extension of generic parametric whose coefficients functions set effect-modifying covariates instead fitted constants. They capable achieving higher model complexity while preserving the structure underlying models, hence generating interpretable predictions. In this paper we study use gradient boosted decision trees as those coefficient-deciding in varying with linearly structured outputs. contrast to traditional choices splines or kernel smoothers, more since they require no structural assumptions effect modifier space. We introduce our proposed method from perspective localized version descent, prove its theoretical consistency under mild commonly adapted by tree research, and empirically demonstrate that achieve high performance qualified their training speed, prediction accuracy intelligibility compared several benchmark algorithms.