Distributed Soft Bayesian Additive Regression Trees

TL;DR

This paper introduces a distributed version of SBART that leverages MPI for scalable computation, enabling efficient Bayesian additive regression modeling on massive datasets and extending its application to classification tasks.

Contribution

The paper presents a modified distributed SBART algorithm using MPI, significantly improving computational speed and scalability for large-scale Bayesian additive modeling, including classification.

Findings

Distributed SBART scales nearly linearly with processor cores.

The approach handles datasets too large for single repositories.

Distributed SBART outperforms traditional BART in classification tasks.

Abstract

Bayesian Additive Regression Trees(BART) is a Bayesian nonparametric approach which has been shown to be competitive with the best modern predictive methods such as random forest and Gradient Boosting Decision Tree.The sum of trees structure combined with a Bayesian inferential framework provide a accurate and robust statistic method.BART variant named SBART using randomized decision trees has been developed and show practical benefits compared to BART. The primary bottleneck of SBART is the speed to compute the sufficient statistics and the publicly avaiable implementation of the SBART algorithm in the R package is very slow.In this paper we show how the SBART algorithm can be modified and computed using single program,multiple data(SPMD) distributed computation with the Message Passing Interface(MPI) library.This approach scales nearly linearly in the number of processor cores, enabling the practitioner to perform statistical inference on massive datasets. Our approach can also handle datasets too massive to fit on any single data repository.We have made modification to this algorithm to make it capable to handle classfication problem which can not be done with the original R package.With data experiments we show the advantage of distributed SBART for classfication problem compared to BART.