GGBoost: Graph-split gradient boosting trees

doi:10.1016/j.neucom.2026.134019

GGBoost: Graph-split gradient boosting trees

2026-05-19

0

PRE

AI

DOI：10.1016/j.neucom.2026.134019

Original

Original request for help

Save

Abstract

En 中文

• We propose a graph-split-based gradient boosting trees method, called GGBoost, for nonparametric regression and classification learning tasks on complex data with graph relations (e.g., spatial data). In GGBoost, a novel graph split decision rule replaces the conventional axis-parallel split rule in XGBoost and RF, allowing for highly flexible, nonlinear decision boundaries that comply with the graph structure. Traditional ensemble tree models can be unified into this general framework. • We propose a novel algorithm that uses recursive gradient updates on rooted and oriented spanning trees for split finding and propose several other computational strategies, such as binning on graphs and row/column subsampling, to ensure the computational efficiency of GGBoost. • We implemented GGBoost in C++ with an interface to R and will include the GitHub repo upon acceptance of the manuscript. The superior performance of the proposed method is demonstrated on extensive simulation and benchmark datasets.

Keywords:

Graph-split

Gradient boosting

Nonparametric regression

Classification

Complex data

GGBoost: Graph-split gradient boosting trees

Abstract

Journal

Organization