R使用xgboost进行交叉验证和调整

示例

library(caret) # for dummyVars
library(RCurl) # download https data
library(Metrics) # calculate errors
library(xgboost) # model

###############################################################################
# Load data from UCI Machine Learning Repository (http://archive.ics.uci.edu/ml/datasets.html)
urlfile <- 'https://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.data'
x <- getURL(urlfile,ssl.verifypeer= FALSE)
adults <- read.csv(textConnection(x), header=F)

# adults <-read.csv('https://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.data', header=F)
names(adults)=c('age','workclass','fnlwgt','education','educationNum',
                'maritalStatus','occupation','relationship','race',
                'sex','capitalGain','capitalLoss','hoursWeek',
                'nativeCountry','income')
# clean up data
adults$income <- ifelse(adults$income==' <=50K',0,1)
# binarize all factors
library(caret)
dmy <- dummyVars(" ~ .", data = adults)
adultsTrsf <- data.frame(predict(dmy, newdata = adults))
###############################################################################

# what we're trying to predict adults that make more than 50k
outcomeName <- c('income')
# list of features
predictors <- names(adultsTrsf)[!names(adultsTrsf) %in% outcomeName]

# play around with settings of xgboost - eXtreme Gradient Boosting (Tree) library
# https://github.com/tqchen/xgboost/wiki/Parameters
#max.depth- maximum depth of the tree
# nrounds - the max number of iterations

# take first 10% of the data only!
trainPortion <- floor(nrow(adultsTrsf)*0.1)

trainSet <- adultsTrsf[ 1:floor(trainPortion/2),]
testSet <- adultsTrsf[(floor(trainPortion/2)+1):trainPortion,]

smallestError <- 100
for (depth in seq(1,10,1)) {
        for (rounds in seq(1,20,1)) {
               
                # train
                bst <- xgboost(data = as.matrix(trainSet[,predictors]),
                               label = trainSet[,outcomeName],
                               max.depth=depth, nround=rounds,
                               objective = "reg:linear", verbose=0)
                gc()
               
                # predict
                predictions <- predict(bst, as.matrix(testSet[,predictors]), outputmargin=TRUE)
                err <- rmse(as.numeric(testSet[,outcomeName]), as.numeric(predictions))
               
                if (err < smallestError) {
                        smallestError = err
                        print(paste(depth,rounds,err))
                }     
        }
} 

cv <- 30
trainSet <- adultsTrsf[1:trainPortion,]
cvDivider <- floor(nrow(trainSet) / (cv+1))

smallestError <- 100
for (depth in seq(1,10,1)) { 
        for (rounds in seq(1,20,1)) {
                totalError <- c()
                indexCount <- 1
                for (cv in seq(1:cv)) {
                        # assign chunk to data test
                        dataTestIndex <- c((cv * cvDivider):(cv * cvDivider + cvDivider))
                        dataTest <- trainSet[dataTestIndex,]
                        # everything else to train
                        dataTrain <- trainSet[-dataTestIndex,]
                       
                        bst <- xgboost(data = as.matrix(dataTrain[,predictors]),
                                       label = dataTrain[,outcomeName],
                                       max.depth=depth, nround=rounds,
                                       objective = "reg:linear", verbose=0)
                        gc()
                        predictions <- predict(bst, as.matrix(dataTest[,predictors]), outputmargin=TRUE)
                       
                        err <- rmse(as.numeric(dataTest[,outcomeName]), as.numeric(predictions))
                        totalError <- c(totalError, err)
                }
                if (mean(totalError) < smallestError) {
                        smallestError = mean(totalError)
                        print(paste(depth,rounds,smallestError))
                } 
        }
} 
 
###########################################################################
# Test both models out on full data set

trainSet <- adultsTrsf[ 1:trainPortion,]

# assign everything else to test
testSet <- adultsTrsf[(trainPortion+1):nrow(adultsTrsf),]

bst <- xgboost(data = as.matrix(trainSet[,predictors]),
               label = trainSet[,outcomeName],
               max.depth=4, nround=19, objective = "reg:linear", verbose=0)
pred <- predict(bst, as.matrix(testSet[,predictors]), outputmargin=TRUE)
rmse(as.numeric(testSet[,outcomeName]), as.numeric(pred))

bst <- xgboost(data = as.matrix(trainSet[,predictors]),
               label = trainSet[,outcomeName],
               max.depth=3, nround=20, objective = "reg:linear", verbose=0)
pred <- predict(bst, as.matrix(testSet[,predictors]), outputmargin=TRUE)
rmse(as.numeric(testSet[,outcomeName]), as.numeric(pred))