README.md

In LTTGH/drkm4mc: Kernel machines with missing covariates

drkm4mc

R/drkm4mc is an R package. This package is used to construct kernel machines with missing covariates for the problem of classification. A doubly robust kernel machine and a weighted-complete-case kernel machies are given.

Installation

You can install it from its GitHub repository. You first need to install the devtools package.

install R/drkm4mc using the install_github function in the devtools package.

library(devtools)
install_github("LTTGH/drkm4mc")

Example use

The example is the veterbral dataset, this dataset is available from the UCI machine learning repository (www.ics.uci.edu/˜mlearn/MLRepository.html) and used in the paper kernel machines for missing covariates. The dataset was split into two sets, the first dataset is used to train the doubly-robust kernel machine with samplesize 200. The remaining dataset is used to test the doubly-robust kernel machine with samplesize 110.

library(drkm4mc)
vData=VertebralData
trainSampleSize=200
trainLable=sample(c(1:310),trainSampleSize,replace = F)
trainData=vData[trainLable,] 
testLabel=setdiff(c(1:310),trainLable)
testData=vData[testLabel,]

First generate the missing mechanism in the training dataset. Assume that x4 and x5 are the covariates that are popentially missing.

alpha=2
beta=-11
stdx1=(trainData$x1-mean(trainData$x1))/sd(trainData$x1)
stdx2=(trainData$x2-mean(trainData$x2))/sd(trainData$x2)
PScov=(stdx1+stdx2)/2
probR=exp(alpha+beta*(PScov*trainData$Y))/(1+exp(alpha+beta*(PScov*trainData$Y)))
R=rbinom(trainSampleSize,1,probR)

Vertebral data with missing covariates is

misCovDat=as.data.frame(cbind(trainData$x1,trainData$x2,trainData$x3,trainData$x6,
                              trainData$x4,trainData$x5,R,trainData$Y,trainData$Y1))

Set the pathes of imputation model and propensity score model

psPath=url(("https://github.com/LTTGH/drkm4mc/raw/master/ModelPS1.RDS"))
impPath=url(("https://github.com/LTTGH/drkm4mc/raw/master/ModelImp1.RDS"))

Train the doubly-robust kernel machines

options(warn=-1)
trainRes=trainingMisCov(misCovDat,px=4,kerType="RBF",kerMethod="DR",lambda=1,sigma=0.1,psPath,impPath,B=2,testPurpose="testing")

Using the doubly-robust kernel machines to predict Y in the testing dataset and show the classification error

dataTst=as.data.frame(cbind(testData$x1,testData$x2,testData$x3,
                        testData$x6,testData$x4,testData$x5,rep(1,110),testData$Y))
1-mean(testingMisCov(trainRes,dataTst)==testData$Y)

LTTGH/drkm4mc documentation built on July 3, 2021, 4:15 p.m.