10 Structuring data transformation and model assessments

We can train the model using Caret package which is an interface that supports many models.

See the XGBoost file.

library(recipes)
library(caret)
library(ISLR)
library(doParallel)

#Initiating parellels
CoreCount  <- makePSOCKcluster(detectCores()-1)
registerDoParallel(CoreCount)

#Loading Data
  set.seed(1337)
  df <- ISLR::OJ
  index <- sample(x = nrow(df),size = 800)
  df.train <- df[index, ]
  df.test <- df[-index, ]  
  
  #Blueprinting - transforming data
    {#Run here
    # blueprint <- {recipe(formula = Purchase ~ . #Defining Y and X's
    #                    ,data = df)} #Selecting the data
    #   #Notice, one could apply steps in the braces 
    #  
    # #blueprint #Could be run to see the operations
    # 
    # 
    # 
    # #Prepping
    # prep.df <- prep(x = blueprint,training = df)
    # #prep.df #Could be run to see the operations
    # 
    # #Baking
    # baked.df <- bake(prep.df, new_data = df)
    # baked.df.train <- bake(prep.df, new_data = df.train) #Used if Blueprint is run on train data
    # baked.df.test <- bake(prep.df, new_data = df.test) #Used if Blueprint is run on train data
    }
  
#Selecting the parameters
train.param <- trainControl(method = "cv" #The resampling method
                            ,number = 5 #Relevant for when the models are trained
                            ,classProbs = TRUE
                            ,summaryFunction = twoClassSummary #Returns sens, spec, ROC
                            ,allowParallel = TRUE
                            )


#Training different models

  #(1) Conditional Inference Tree
  model.tree <- train(Purchase ~ .
                      ,data = df.train
                      ,method = "ctree2" #See list http://topepo.github.io/caret/train-models-by-tag.html
                      ,trControl = train.param
                      ,metric = "ROC")
  
  #(2) Random Forest
  model.rf.1 <- train(Purchase ~ .
                      ,data = df.train
                      ,method = "rf"
                      ,ntree = 10 #Number of trees to grow
                      ,trControl = train.param
                      ,metric = "ROC")
  
  #(3) Random Forest 2
  model.rf.2 <- train(Purchase ~ .
                      ,data = df.train
                      ,method = "rf"
                      ,ntree = 150 #Number of trees to grow
                      ,trControl = train.param
                      ,metric = "ROC")

  #(4) XGBoosting
  model.xgboost <- train(Purchase ~ .
                         ,data = df.train
                         ,method = "xgbTree"
                         ,trControl = train.param
                         ,metric = "ROC"
                         ,tuneGrid = data.frame(.nrounds = 300 #Defining a table of tuning params
                                                ,.max_depth = 3
                                                ,.eta = 0.03
                                                ,.gamma = 0
                                                ,.subsample = 0.5
                                                ,.colsample_bytree = 0.1
                                                ,.min_child_weight = 1))
  
#Terminating the parallels
stopCluster(CoreCount)
registerDoSEQ()

#Summarizing the results
results <- resamples(list(tree = model.tree,
                          rf.1  = model.rf.1,
                          rf.2  = model.rf.2,
                          xgboost = model.xgboost)
                     )
results

summary(results)
bwplot(results)
dotplot(results)