Random Forest

LOADING DATA INTO R ENVIRONMENT

Reading the Data

# reading the data
library(data.table)
df <- fread("MCICreditCardDefault.csv")

Converting some data columns into factors

# convert 'Male' as a factor
df[, Male := as.factor(Male)]
# convert 'Education' as a factor
df[, Education := as.factor(Education)]
# convert 'MaritalStatus' as a factor
df[, MaritalStatus := as.factor(MaritalStatus)]
# convert 'Default' as a factor
df[, Default := as.factor(Default)]

Splitting data into training and test set

library(caret)
# data partition
set.seed(2341)
trainIndex <- createDataPartition(df$Default, p = 0.80, list = FALSE)
# 80% training data
train.df <- df[trainIndex, ]
# 20% testing data
test.df <- df[-trainIndex, ]
dim(train.df)

## [1] 23681     9

dim(test.df)

## [1] 5920    9

TRAINING THE RANDOM FOREST MODEL

Setting Control parameters

library(caret)
# control parameters
trctrl <- trainControl(method = "none", classProbs = TRUE,)

Running the Training Model

## setting levels as "Yes" and "NO"
train.df$Default <- ifelse(train.df$Default == "1","Yes","No")

set.seed(123)
# fitting decision tree classification model
ModelRF <- train(Default ~ CreditLimit
                         + Male
                         + Education
                         + MaritalStatus
                         + Age
                         + BillOutstanding
                         + LastPayment, 
                         data = train.df, 
                         method = "rf",
                         nbagg = 50,
                         parms  = list(split = "gini"),
                         trControl = trctrl,
                         importance = TRUE)
# model summary
ModelRF

## Random Forest 
## 
## 23681 samples
##     7 predictor
##     2 classes: 'No', 'Yes' 
## 
## No pre-processing
## Resampling: None

Variable Importance in Random Forest Model

# plotting variable importance
plot(varImp(ModelRF))

TESTING THE RANDOM FOREST MODEL

Predicting Model on Test Data Set

# predicting the model on test data set
PredRFModel <- predict(ModelRF, 
                    newdata = test.df,type = "prob")

Plotting the Predicted Probabilities

# plot of probabilities
plot(PredRFModel$Yes, 
     main = "Scatterplot of Probabilities of default (test data)", 
     xlab = "Customer ID", 
     ylab = "Predicted Probability of default")

Confusion Matrix at 50% Cut-Off Probability

# taking the cut-off probability 50%
Predicted <- ifelse(PredRFModel$Yes > 0.50, "Yes", "No")

# ordering the vectors
Actual <- ifelse(test.df$Default=="1","Yes","No")
# making confusion matrix
cm <-confusionMatrix(table(Predicted,Actual), positive = "Yes")
cm

## Confusion Matrix and Statistics
## 
##          Actual
## Predicted   No  Yes
##       No  4519 1226
##       Yes   80   95
##                                           
##                Accuracy : 0.7794          
##                  95% CI : (0.7686, 0.7899)
##     No Information Rate : 0.7769          
##     P-Value [Acc > NIR] : 0.3262          
##                                           
##                   Kappa : 0.0789          
##                                           
##  Mcnemar's Test P-Value : <2e-16          
##                                           
##             Sensitivity : 0.07192         
##             Specificity : 0.98260         
##          Pos Pred Value : 0.54286         
##          Neg Pred Value : 0.78660         
##              Prevalence : 0.22314         
##          Detection Rate : 0.01605         
##    Detection Prevalence : 0.02956         
##       Balanced Accuracy : 0.52726         
##                                           
##        'Positive' Class : Yes             
## 

ROC Curve using Package ROCR

# loading the package
library(ROCR)
RFPrediction <- predict(ModelRF, test.df,type = "prob")
Prediction <- prediction(RFPrediction[2],test.df$Default)
performance <- performance(Prediction, "tpr","fpr")
# plotting ROC curve
plot(performance,main = "ROC Curve",col = 2,lwd = 2)
abline(a = 0,b = 1,lwd = 2,lty = 3,col = "black")

AUC (Area Under the Curve)

library(ROCR)
# area under curve
RFPrediction <- prediction(PredRFModel[2],test.df$Default)
aucRF <- performance(RFPrediction, measure = "auc")
aucRF <- aucRF@y.values[[1]]
aucRF

## [1] 0.6459245