punoqun / GBDTMO

Gradient boosted decision trees for multiple outputs. Better generalization ability, faster training and inference.

GBDT-MO: Gradient Boosted Decision Tree for Multiple Outputs

Important Updates

• 2021.03.09: the implementation for updating multi-output histograms is problematic. Now it has been corrected. Please check src/dataStruct.h for details. This improves the convergence speed. Experiment results are slightly different (generally better) than the old one.

Introduction

Gradient boosted decision tree (GBDT) is a popular machine learning algorithm. Current open-sourced GBDT implementations are mainly designed for single output. When there are multiple outputs, they build multiple trees each of which corresponds to an output variable. Such a strategy ignores the correlations between output variables which leads to worse generalization ability and tree redundancy.

To address this problem, we propose a general method to learn GBDT for multiple outputs (GBDT-MO). Each leaf of GBDT-MO constructs the predictions of all variables or a subset of automatically selected variables. Experiments show that GBDT-MO has better generalization ability and faster training and inference speed than GBDT for single output.

GBDT-MO suffers from slow convergence (requires more boost rounds). It should be improved in the future.

Implementations

We implement GBDT-MO from scratch by C++. And we provide a Python interface. Our implementations are similar to LightGBM except the learning mechanisms designed for multiple outputs. Some advanced features are not included, such as GPU training and distributed training. This project is only tested on Linux and used for academic explorations currently.

Installation

First, clone this project and compile the C++ source code:

bash make.sh

The shared library will be generated in:

build/gbdtmo.so

Then, install the Python package:

pip install .

Results

We show test performance on six real-world datasets. Here, GBDT-SO is our own implementation of GBDT for single output.

Dataset	MNIST	Yeast	Caltech101	NUS-WIDE	MNIST-inpaining
Metric	accuracy	accuracy	accuracy	top-1 accuracy	RMSE
XGBoost	97.86	62.94	56.52	43.72	0.26088
LightGBM	98.03	61.97	55.94	43.99	0.26090
GBDT-sparse	96.41	62.83	43.93	44.05	-
GBDT-SO	98.08	61.97	56.62	44.10	0.26157
GBDT-MO	98.30	62.29	57.49	44.21	0.26025

We show training time averaged by rounds in log scale.

We show inference time on test set averaged by rounds (millisecond).

Dataset	MNIST	Caltech101	NUS-WIDE	MNIST-inpaining
XGBoost	6.38	9.57	119.24	6.61
LightGBM	1.78	2.76	209.79	8.50
GBDT-SO	3.77	4.70	515.90	12.75
GBDT-MO	0.59	0.15	7.83	0.32

Links

• For algorithm and experiment details, refer our paper.
• For examples or reproducing the results, refer GBDTMO-EX.
• For instructions, refer our documentation.