## xgboost using Caret
#library(caret)
pacman::p_load(doMC)
registerDoMC(cores=4)
options(scipen = 999)

pacman::p_load(mlbench)
data("Glass", package = "mlbench")

unique(Glass$Type)

levels(Glass$Type)<- c(0:5) # propoer sequence

# For multiclass classification problem the response should be numbers starting from zero

Glass$Type <- as.integer(as.character(Glass$Type))
unique(Glass$Type)
colnames(Glass)

pacman::p_load(xgboost)
pacman::p_load(caret)

set.seed(100)

trainRows <- createDataPartition(Glass$Type,p=.7,list=FALSE)

trainData <- Glass[trainRows,]
testData <- Glass[-trainRows,]

'%ni%' <- Negate('%in%')

# creating the matrix for training the model

trainData_xg <- xgb.DMatrix(data.matrix(trainData[,colnames(trainData) %ni% 'Type']),
                            label=as.numeric(trainData$Type))

testData_xg <- xgb.DMatrix(data.matrix(testData[,colnames(testData) %ni% 'Type']),
                           label=as.numeric(testData$Type))

watchlist <- list(train = trainData_xg, test = testData_xg)



# Define the parameters and cross validate
param <- list("objective"="multi:softprob",
              "eval_metric"="mlogloss",
              #watchlist=watchlist,
              num_class=length(unique(trainData$Type)))

cv.nround <- 5
cv.nfold <- 3

cvMod <- xgb.cv(param=param,
                data=trainData_xg,
                nfold=cv.nfold,
                nrounds=cv.nround,
                watchlist=watchlist)
cvMod

# Build the XGBoost model

nrounds <- 5000

xgMod <- xgboost(param=param,data = trainData_xg, nrounds=nrounds)
xgMod2 <- xgb.train(param=param,data = trainData_xg, nrounds=nrounds,watchlist = watchlist)
# See the xgBoost Model
model <- xgb.dump(xgMod,with_stats = T)
model[1:10]

model2 <- xgb.dump(xgMod2,with_stats = T)
model2[1:10]

## get the feature important
#
names <- colnames(trainData)[colnames(trainData) %ni% 'Type']

# compute feature importance matrix
featureImp <- xgb.importance(names, model=xgMod)
featureImp

featureImp2 <- xgb.importance(names, model=xgMod2)
featureImp2
# graph
xgb.plot.importance(featureImp)
xgb.plot.importance(featureImp2)
# Plot
pacman::p_load(XML)
pacman::p_load(DiagrammeR)

xgb.plot.tree(feature_names = names,
              model=xgMod, n_first_tree = 2)
xgb.plot.tree(feature_names = names,
              model=xgMod2, n_first_tree = 2)

# Predict
pred <- predict(xgMod, testData_xg)
pred2 <- predict(xgMod2, testData_xg)

# MAPE

tab <- table(pred,trainData$Type)
tab <- table(pred2,trainData$Type)

caret::confusionMatrix(tab)