A set of programms implementing algorithms on Subgroup Discovery under Treatment Effect. The name of the library come from SUbgroup DIscovery under Treatment Effect. Apparently in Russian the word SuDiTE is very close to "you judge". It is close to the goal of the library which is deciding which algorithm is the best one.
A special package in R is created for testing and comparing of different algos.
In order to install the package one is need to install 'devtools' library.
install.packages("devtools")
Then our package can be installed from GitHub
devtools::install_github("AlekseyBuzmakov/SuDiTE", subdir = "SuDiTE.R")
The following functions are provided by the package
evaluateAlgos(trainModelFuncs, predictModelFuncts, subgroupQualityFunc, trainY, trainTrt, TrainX, testY, testTrt, testX)
The first two arguments are a vector of functions of the same length. The first ones train models, and the second ones predict with the corresponding models.
subgroupQualityFunc is the function that measure the quality of the subgroup. The simplest case as Average Treatment Effect is given by function SuDiTE::subgroupAverageTreatmentEffect()
train* are response, treatment and covariates of the training dataset test* are response, treatment and covariates of the test dataset
The result is the quality measured in the test dataset for the models build in the train dataset.
crossValidateAlgos(trainModelFuncs, predictModelFuncs, subgroupQualityFunc, dbY, dbTrt, dbX, numTrials, testProportion)
The first three arguments are the same as for SuDiTE::evaluateAlgos function. db* is response, treatment and covariates for the dataset. numTrials is the number of divisions to train-hold_out datasets testProportion is the 0-1 number specifying the proportion of the whole dataset to be considered as the hold-out set
subgroupQualityFunc(subgroup, Y, Trt)
Function measure the quality of a subgroup subgroup is a vector of logical flags specifying if i-th object is included into the subgroup Y, Trt is the value of response and treatment varibale for every object