partDSA: partDSA

In AMolinaro/partDSA-Code: Partitioning Using Deletion, Substitution, and Addition Moves

Description

partDSA is a novel tool for generating a piecewise constant estimation sieve of candidate estimators based on an intensive and comprehensive search over the entire covariate space. The strength of this algorithm is that it builds 'and' and 'or' statements. This allows combinations and substitutions of regions for the purpose of discovering intricate correlations patterns and interactions in addition to main effects. Depending on the application, this approach will supersede methods such as CART by being not only more aggressive but also more flexible. As such, partDSA provides the user an additional tool for their statistical toolbox.

Usage

	
partDSA(x, y, wt=rep(1, nrow(x)), x.test=x, y.test=y, wt.test,
        control=DSA.control(), sleigh)
DSA.control(vfold=10, minsplit = 20, minbuck=round(minsplit/3), 
                        cut.off.growth=10, MPD=0.1, missing="impute.at.split", 
                        loss.function="default", wt.method="KM", brier.vec=NULL,
                        leafy=0, leafy.random.num.variables.per.split=4,
                        leafy.num.trees=50, leafy.subsample=0, save.input=FALSE, 
                        boost=0, boost.rounds=100, cox.vec=NULL,IBS.wt=NULL, partial=NULL)

Arguments

x	The matrix or data frame of predictor variables for the training set, used to build the model. Each row corresponds to an observation, and each column corresponds to a variable.
y	The outcome (response) vector, either continuous or categorical, representing the true response values for observations in x. The length of this vector should equal the number of rows in x.
wt	Optional vector of training weights with length equal to the number of observations in x. Default is a vector of ones with length equal to the number of training set observations.
x.test	The matrix or data frame of predictor variables used to build the model. The number of columns (variables) of x.test should equal the number of columns as x. The default is x.
y.test	The outcome (response) vector, either continuous or categorical, representing the true response values for observations in x.test. The length of this vector should equal the number of rows in x.test. The default value is y.
wt.test	Optional vector of test weights with length equal to the number of test set observations. Default value is wt if x.test wasn't specified, otherwise it is a vector of ones with length equal to the number of test set observations.
control	A list object used to specify additional control parameters. This is normally created by calling the DSA.control function. Default value is the result of calling DSA.control with no arguments.
sleigh	Optional sleigh object to allow the cross-validation to be performed in parallel using the nws package. If not specified, the cross-validation will be executed sequentially.
vfold	The number of folds of cross-validation for the model building process. The default value is 10.
minsplit	The minimum number of observations in order to split a partition into two paritions. The default value is 20.
minbuck	The minimum number of observations in any terminal partition. The default value is round(minsplit/3).
cut.off.growth	The maximum number of terminal partitions to be considered when building the model. The default value is 10.
MPD	Minimum Percent Difference. The model fit must improve by this percentage in order to be considered. This saves time in the model building process. The default value is 0.1.
missing	Character string specifying how missing data should be handled. The default value is "no." See the details section from more information.
loss.function	The function to be minimized when building the model. For categorical outcomes, "entropy" (default) or "gini" can be specified. For continuous outcomes, the L2 loss function is used.
wt.method	Not documented yet.
brier.vec	Not documented yet.
cox.vec	Not documented yet.
IBS.wt	Not documented yet.
leafy	Set to 1 to run Bagged partDSA.
leafy.random.num.variables.per.split	Number of variables to use if utilizing random variable selection in Bagged partDSA.
leafy.num.trees	Numbed in trees in Bagged partDSA. Default is 50.
leafy.subsample	Numeric value between 0 and 1. The value 0 is used for bootstrap sampling (sampling witht replacement). If the value is greater than 0, it corresponds to the proportion of samples used to build the model, such as 0.632. The default is 0.
save.input	Indicates if x and y should be saved in the object returned by partDSA. If FALSE, x and y are set to NULL. The default value is FALSE.
boost	Set to 1 to run Boosted partDSA.
boost.rounds	Maximum number of rounds of boosting. Default is 100.
partial	If set to "deciles," step partial importance is computed on deciles of data rather than actual data.

Details

missing set to "no" indicates that there is no missing data and will create an error if missing data is found in the dataset. Setting missing="impute.at.split" will use a data imputation method similar to that in CRUISE (Kim and Loh, 2001). At each split, the non-missing observations for a given variable will be used to find the best split, and the missing observations will be imputed based on the mean or mode (depending on whether the variable is categorical or continuous) of the non-missing observations in that node. Once the node assignment of these missing observations is determined using the imputed values, the imputed values are returned to their missing status. For missing values in the test set, the grand mean or mode from the corresponding variables in the training set are used. Including variables which are entirely missing will result in an error.

Examples

library(MASS)
set.seed(6442)

n <- nrow(Boston)
tr.n <- floor(n / 2)
train.index <- sample(1:n, tr.n, replace=FALSE)
test.index <- (1:n)[-train.index]

x <- Boston[train.index, -14]
y <- Boston[train.index, 14]
x.test <- Boston[test.index, -14]
y.test <- Boston[test.index, 14]

control <- DSA.control(vfold=1)  # no cross-validation
partDSA(x, y, x.test=x.test, y.test=y.test, control=control)

AMolinaro/partDSA-Code documentation built on May 5, 2019, 11:38 a.m.