BART is a Bayesian “sum-of-trees” model.
For a numeric response \(y\), we have
\(y = f(x) + \epsilon\),
where \(\epsilon \sim N(0,\sigma^2)\).
\(f\) is the sum of many tree models. The goal is to have very flexible inference for the uknown function \(f\).
In the spirit of “ensemble models”, each tree is constrained by a prior to be a weak learner so that it contributes a small amount to the overall fit.
gbart(
x.train, y.train,
x.test=matrix(0,0,0), type='wbart',
ntype=as.integer(
factor(type, levels=c('wbart', 'pbart', 'lbart'))),
sparse=FALSE, theta=0, omega=1,
a=0.5, b=1, augment=FALSE, rho=NULL,
xinfo=matrix(0,0,0), usequants=FALSE,
rm.const=TRUE,
sigest=NA, sigdf=3, sigquant=0.90,
k=2, power=2, base=0.95,
lambda=NA, tau.num=c(NA, 3, 6)[ntype],
offset=NULL, w=rep(1, length(y.train)),
ntree=c(200L, 50L, 50L)[ntype], numcut=100L,
ndpost=1000L, nskip=100L,
keepevery=c(1L, 10L, 10L)[ntype],
printevery=100L, transposed=FALSE,
hostname=FALSE,
mc.cores = 1L, ## mc.gbart only
nice = 19L, ## mc.gbart only
seed = 99L ## mc.gbart only
)mc.gbart(
x.train, y.train,
x.test=matrix(0,0,0), type='wbart',
ntype=as.integer(
factor(type, levels=c('wbart', 'pbart', 'lbart'))),
sparse=FALSE, theta=0, omega=1,
a=0.5, b=1, augment=FALSE, rho=NULL,
xinfo=matrix(0,0,0), usequants=FALSE,
rm.const=TRUE,
sigest=NA, sigdf=3, sigquant=0.90,
k=2, power=2, base=0.95,
lambda=NA, tau.num=c(NA, 3, 6)[ntype],
offset=NULL, w=rep(1, length(y.train)),
ntree=c(200L, 50L, 50L)[ntype], numcut=100L,
ndpost=1000L, nskip=100L,
keepevery=c(1L, 10L, 10L)[ntype],
printevery=100L, transposed=FALSE,
hostname=FALSE,
mc.cores = 2L, nice = 19L, seed = 99L
)
gbart returns an object of type gbart which is
essentially a list.
In the numeric \(y\) case, the list has components:
A matrix with ndpost rows and nrow(x.train) columns.
Each row corresponds to a draw \(f^*\) from the posterior of \(f\)
and each column corresponds to a row of x.train.
The \((i,j)\) value is \(f^*(x)\) for the \(i^{th}\) kept draw of \(f\)
and the \(j^{th}\) row of x.train.
Burn-in is dropped.
Same as yhat.train but now the x's are the rows of the test data.
train data fits = mean of yhat.train columns.
test data fits = mean of yhat.test columns.
post burn in draws of sigma, length = ndpost.
burn-in draws of sigma.
a matrix with ndpost rows and nrow(x.train) columns. Each row is for a draw. For each variable (corresponding to the columns), the total count of the number of times that variable is used in a tree decision rule (over all trees) is given.
The rough error standard deviation (\(\sigma\)) used in the prior.
Explanatory variables for training (in sample)
data.
May be a matrix or a data frame, with (as usual) rows
corresponding to observations and columns to variables.
If a
variable is a factor in a data frame, it is replaced with dummies.
Note that \(q\) dummies are created if \(q>2\) and one dummy
created if \(q=2\) where \(q\) is the number of levels of the
factor. gbart will generate draws of \(f(x)\) for each
\(x\) which is a row of x.train.
Continuous or binary dependent variable for training (in sample) data.
If \(y\) is numeric, then a continuous BART model is fit (Normal errors).
If \(y\) is binary (has only 0's and 1's), then a binary BART model
with a probit link is fit by default: you can over-ride the default via the
argument type to specify a logit BART model.
Explanatory variables for test (out of sample)
data. Should have same structure as x.train.
gbart will generate draws of \(f(x)\) for each \(x\) which
is a row of x.test.
You can use this argument to specify the type of fit.
'wbart' for continuous BART, 'pbart' for probit BART or
'lbart' for logit BART.
The integer equivalent of type where
'wbart' is 1, 'pbart' is 2 and
'lbart' is 3.
Whether to perform variable selection based on a sparse Dirichlet prior rather than simply uniform; see Linero 2016.
Set \(theta\) parameter; zero means random.
Set \(omega\) parameter; zero means random.
Sparse parameter for \(Beta(a, b)\) prior: \(0.5<=a<=1\) where lower values inducing more sparsity.
Sparse parameter for \(Beta(a, b)\) prior; typically, \(b=1\).
Sparse parameter: typically \(rho=p\) where \(p\) is the number of covariates under consideration.
Whether data augmentation is to be performed in sparse variable selection.
You can provide the cutpoints to BART or let BART
choose them for you. To provide them, use the xinfo
argument to specify a list (matrix) where the items (rows) are the
covariates and the contents of the items (columns) are the
cutpoints.
If usequants=FALSE, then the
cutpoints in xinfo are generated uniformly; otherwise,
if TRUE, uniform quantiles are used for the cutpoints.
Whether or not to remove constant variables.
The prior for the error variance
(\(sigma^2\)) is inverted chi-squared (the standard
conditionally conjugate prior). The prior is specified by choosing
the degrees of freedom, a rough estimate of the corresponding
standard deviation and a quantile to put this rough estimate at. If
sigest=NA then the rough estimate will be the usual least squares
estimator. Otherwise the supplied value will be used.
Not used if \(y\) is binary.
Degrees of freedom for error variance prior. Not used if \(y\) is binary.
The quantile of the prior that the rough estimate
(see sigest) is placed at. The closer the quantile is to 1, the more
aggresive the fit will be as you are putting more prior weight on
error standard deviations (\(sigma\)) less than the rough
estimate. Not used if \(y\) is binary.
For numeric \(y\), k is the number of prior
standard deviations \(E(Y|x) = f(x)\) is away from +/-0.5. For
binary \(y\), k is the number of prior standard deviations
\(f(x)\) is away from +/-3. The bigger k is, the more
conservative the fitting will be.
Power parameter for tree prior.
Base parameter for tree prior.
The scale of the prior for the variance. If lambda is zero,
then the variance is to be considered fixed and known at the given
value of sigest. Not used if \(y\) is binary.
The numerator in the tau definition, i.e.,
tau=tau.num/(k*sqrt(ntree)).
Continous BART operates on y.train centered by
offset which defaults to mean(y.train). With binary
BART, the centering is \(P(Y=1 | x) = F(f(x) + offset)\) where
offset defaults to F^{-1}(mean(y.train)). You can use
the offset parameter to over-ride these defaults.
Vector of weights which multiply the standard deviation. Not used if \(y\) is binary.
The number of trees in the sum.
The number of possible values of \(c\) (see
usequants). If a single number if given, this is used for all
variables. Otherwise a vector with length equal to
ncol(x.train) is required, where the \(i^{th}\)
element gives the number of \(c\) used for the \(i^{th}\)
variable in x.train. If usequants is false, numcut equally
spaced cutoffs are used covering the range of values in the
corresponding column of x.train. If usequants is true, then
\(min(numcut, the number of unique values in the corresponding
columns of x.train - 1)\) values are used.
The number of posterior draws returned.
Number of MCMC iterations to be treated as burn in.
As the MCMC runs, a message is printed every printevery draws.
Every keepevery draw is kept to be returned to the user.
When running gbart in parallel, it is more memory-efficient
to transpose x.train and x.test, if any, prior to
calling mc.gbart.
When running on a cluster occasionally it is useful
to track on which node each chain is running; to do so
set this argument to TRUE.
Setting the seed required for reproducible MCMC.
Number of cores to employ in parallel.
Set the job niceness. The default niceness is 19: niceness goes from 0 (highest) to 19 (lowest).
BART is a Bayesian MCMC method. At each MCMC interation, we produce a draw from the joint posterior \((f,\sigma) | (x,y)\) in the numeric \(y\) case and just \(f\) in the binary \(y\) case.
Thus, unlike a lot of other modelling methods in R, we do not produce
a single model object from which fits and summaries may be extracted.
The output consists of values \(f^*(x)\) (and
\(\sigma^*\) in the numeric case) where * denotes a
particular draw. The \(x\) is either a row from the training data,
x.train or the test data, x.test.
For x.train/x.test with missing data elements, gbart
will singly impute them with hot decking. For one or more missing
covariates, record-level hot-decking imputation deWaPann11 is
employed that is biased towards the null, i.e., nonmissing values
from another record are randomly selected regardless of the
outcome. Since mc.gbart runs multiple gbart threads in
parallel, mc.gbart performs multiple imputation with hot
decking, i.e., a separate imputation for each thread. This
record-level hot-decking imputation is biased towards the null, i.e.,
nonmissing values from another record are randomly selected
regardless of y.train.
pbart
##simulate data (example from Friedman MARS paper)
f = function(x){
10*sin(pi*x[,1]*x[,2]) + 20*(x[,3]-.5)^2+10*x[,4]+5*x[,5]
}
sigma = 1.0 #y = f(x) + sigma*z , z~N(0,1)
n = 100 #number of observations
set.seed(99)
x=matrix(runif(n*10),n,10) #10 variables, only first 5 matter
Ey = f(x)
y=Ey+sigma*rnorm(n)
lmFit = lm(y~.,data.frame(x,y)) #compare lm fit to BART later
##test BART with token run to ensure installation works
set.seed(99)
bartFit = wbart(x,y,nskip=5,ndpost=5)
if (FALSE) {
##run BART
set.seed(99)
bartFit = wbart(x,y)
##compare BART fit to linear matter and truth = Ey
fitmat = cbind(y,Ey,lmFit$fitted,bartFit$yhat.train.mean)
colnames(fitmat) = c('y','Ey','lm','bart')
print(cor(fitmat))
}
Run the code above in your browser using DataLab