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.
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")
install.packages("bark", repos = c('https://merliseclyde.r-universe.dev', 'https://cloud.r-project.org'))
(verify that the branch has a passing R CMD check badge above)
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.
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!