bark: Bayesian Additive Regression Kernels

The bark package implements estimation for a Bayesian nonparametric regression model represented as a sum of multivariate Gaussian kernels as a flexible model to capture nonlinearities, interactions and feature selection.

Installation

You can install the released version of bark from CRAN with:

install.packages("bark")

And the development version from GitHub with:

require("devtools")
devtools::install_github("merliseclyde/bark")

(verify that the branch has a passing R CMD check badge above)

Example

library(bark)
set.seed(42)
traindata <- sim_Friedman2(200, sd=125)
testdata <- sim_Friedman2(1000, sd=0)
fit.bark.d <- bark(y ~ .,  
                   data=data.frame(traindata), 
                   testdata = data.frame(testdata),
                   classification=FALSE, 
                   selection = TRUE,
                   common_lambdas = FALSE,
                   printevery = 10^10)

mean((fit.bark.d$yhat.test.mean-testdata$y)^2)
#> [1] 1920.283

bark is similar to SVM, however it allows different kernel smoothing parameters for every dimension of the inputs $x$ as well as selection of inputs by allowing the kernel smoothing parameters to be zero.

The plot below shows posterior draws of the $\lambda$ for the simulated data.

boxplot(as.data.frame(fit.bark.d$theta.lambda))

The posterior distribution for $\lambda_1$ and $\lambda_4$ are concentrated near zero, which leads to $x_1$ and $x_2$ dropping from the mean function.

Roadmap for Future Enhancements

Over the next year the following enhancements are planned:

• port more of the R code to C/C++ for improvements in speed

• add S3 methods for predict, summary, plot

• add additional kernels and LARK methods from AOS (2011) paper

• better hyperparameter specification

If there are features you would like to see added, please feel free to create an issue in GitHub and we can discuss!