Predict on the Risky Vs Good Customer on Fraud check data and its attributes using Random Forest.
#install.packages("randomForest")
#install.packages("Mass")
#install.packages("caret")
library(randomForest)
## Warning: package 'randomForest' was built under R version 3.5.1
## randomForest 4.6-14
## Type rfNews() to see new features/changes/bug fixes.
library(MASS)
library(caret)
## Warning: package 'caret' was built under R version 3.5.1
## Loading required package: lattice
## Loading required package: ggplot2
## Warning: package 'ggplot2' was built under R version 3.5.1
##
## Attaching package: 'ggplot2'
## The following object is masked from 'package:randomForest':
##
## margin
# USe the set.seed function so that we get same results each time
set.seed(123)
FraudCheck <- read.csv(file.choose())
hist(FraudCheck$Taxable.Income)
hist(FraudCheck$Taxable.Income, main = "Sales of Companydata",xlim = c(0,100000),
breaks=c(seq(40,60,80)), col = c("blue","red", "green","violet"))
Risky_Good = ifelse(FraudCheck$Taxable.Income<= 30000, "Risky", "Good")
# if Taxable Income is less than or equal to 30000 then Risky else Good.
FCtemp= data.frame(FraudCheck,Risky_Good)
FC = FCtemp[,c(1:7)]
str(FC)
## 'data.frame': 600 obs. of 7 variables:
## $ Undergrad : Factor w/ 2 levels "NO","YES": 1 2 1 2 1 1 1 2 1 2 ...
## $ Marital.Status : Factor w/ 3 levels "Divorced","Married",..: 3 1 2 3 2 1 1 3 3 1 ...
## $ Taxable.Income : int 68833 33700 36925 50190 81002 33329 83357 62774 83519 98152 ...
## $ City.Population: int 50047 134075 160205 193264 27533 116382 80890 131253 102481 155482 ...
## $ Work.Experience: int 10 18 30 15 28 0 8 3 12 4 ...
## $ Urban : Factor w/ 2 levels "NO","YES": 2 2 2 2 1 1 2 2 2 2 ...
## $ Risky_Good : Factor w/ 2 levels "Good","Risky": 1 1 1 1 1 1 1 1 1 1 ...
table(FC$Risky_Good) # 476 good customers and 124 risky customers
##
## Good Risky
## 476 124
# Data Partition
set.seed(123)
ind <- sample(2, nrow(FC), replace = TRUE, prob = c(0.7,0.3))
train <- FC[ind==1,]
test <- FC[ind==2,]
set.seed(213)
rf <- randomForest(Risky_Good~., data=train)
rf # Description of the random forest with no of trees, mtry = no of variables for splitting
##
## Call:
## randomForest(formula = Risky_Good ~ ., data = train)
## Type of random forest: classification
## Number of trees: 500
## No. of variables tried at each split: 2
##
## OOB estimate of error rate: 0.47%
## Confusion matrix:
## Good Risky class.error
## Good 343 1 0.002906977
## Risky 1 79 0.012500000
# each tree node.
# Out of bag estimate of error rate is 0.47 % in Random Forest Model.
attributes(rf)
## $names
## [1] "call" "type" "predicted"
## [4] "err.rate" "confusion" "votes"
## [7] "oob.times" "classes" "importance"
## [10] "importanceSD" "localImportance" "proximity"
## [13] "ntree" "mtry" "forest"
## [16] "y" "test" "inbag"
## [19] "terms"
##
## $class
## [1] "randomForest.formula" "randomForest"
# Prediction and Confusion Matrix - Training data
pred1 <- predict(rf, train)
head(pred1)
## 1 3 6 7 9 10
## Good Good Good Good Good Good
## Levels: Good Risky
head(train$Risky_Good)
## [1] Good Good Good Good Good Good
## Levels: Good Risky
# looks like the first six predicted value and original value matches.
confusionMatrix(pred1, train$Risky_Good) # 100 % accuracy on training data
## Confusion Matrix and Statistics
##
## Reference
## Prediction Good Risky
## Good 344 0
## Risky 0 80
##
## Accuracy : 1
## 95% CI : (0.9913, 1)
## No Information Rate : 0.8113
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 1
## Mcnemar's Test P-Value : NA
##
## Sensitivity : 1.0000
## Specificity : 1.0000
## Pos Pred Value : 1.0000
## Neg Pred Value : 1.0000
## Prevalence : 0.8113
## Detection Rate : 0.8113
## Detection Prevalence : 0.8113
## Balanced Accuracy : 1.0000
##
## 'Positive' Class : Good
##
# more than 98% Confidence Interval.
# Sensitivity for Yes and No is 100 %
# Prediction with test data - Test Data
pred2 <- predict(rf, test)
confusionMatrix(pred2, test$Risky_Good) # 100 % accuracy on test data
## Confusion Matrix and Statistics
##
## Reference
## Prediction Good Risky
## Good 132 0
## Risky 0 44
##
## Accuracy : 1
## 95% CI : (0.9793, 1)
## No Information Rate : 0.75
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 1
## Mcnemar's Test P-Value : NA
##
## Sensitivity : 1.00
## Specificity : 1.00
## Pos Pred Value : 1.00
## Neg Pred Value : 1.00
## Prevalence : 0.75
## Detection Rate : 0.75
## Detection Prevalence : 0.75
## Balanced Accuracy : 1.00
##
## 'Positive' Class : Good
##
# Error Rate in Random Forest Model :
plot(rf)
# at 200 there is a constant line and it doesnot vary after 200 trees
# Tune Random Forest Model mtry
tune <- tuneRF(train[,-6], train[,6], stepFactor = 0.5, plot = TRUE, ntreeTry = 300,
trace = TRUE, improve = 0.05)
## mtry = 2 OOB error = 43.16%
## Searching left ...
## mtry = 4 OOB error = 44.1%
## -0.02185792 0.05
## Searching right ...
## mtry = 1 OOB error = 43.16%
## 0 0.05
rf1 <- randomForest(Risky_Good~., data=train, ntree = 200, mtry = 2, importance = TRUE,
proximity = TRUE)
rf1
##
## Call:
## randomForest(formula = Risky_Good ~ ., data = train, ntree = 200, mtry = 2, importance = TRUE, proximity = TRUE)
## Type of random forest: classification
## Number of trees: 200
## No. of variables tried at each split: 2
##
## OOB estimate of error rate: 0.24%
## Confusion matrix:
## Good Risky class.error
## Good 343 1 0.002906977
## Risky 0 80 0.000000000
pred1 <- predict(rf1, train)
confusionMatrix(pred1, train$Risky_Good) # 100 % accuracy on training data
## Confusion Matrix and Statistics
##
## Reference
## Prediction Good Risky
## Good 344 0
## Risky 0 80
##
## Accuracy : 1
## 95% CI : (0.9913, 1)
## No Information Rate : 0.8113
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 1
## Mcnemar's Test P-Value : NA
##
## Sensitivity : 1.0000
## Specificity : 1.0000
## Pos Pred Value : 1.0000
## Neg Pred Value : 1.0000
## Prevalence : 0.8113
## Detection Rate : 0.8113
## Detection Prevalence : 0.8113
## Balanced Accuracy : 1.0000
##
## 'Positive' Class : Good
##
# Around 99% Confidence Interval.
# Sensitivity for Yes and No is 100 %
# test data prediction using the Tuned RF1 model
pred2 <- predict(rf1, test)
confusionMatrix(pred2, test$Risky_Good) # 100 % accuracy on test data
## Confusion Matrix and Statistics
##
## Reference
## Prediction Good Risky
## Good 132 0
## Risky 0 44
##
## Accuracy : 1
## 95% CI : (0.9793, 1)
## No Information Rate : 0.75
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 1
## Mcnemar's Test P-Value : NA
##
## Sensitivity : 1.00
## Specificity : 1.00
## Pos Pred Value : 1.00
## Neg Pred Value : 1.00
## Prevalence : 0.75
## Detection Rate : 0.75
## Detection Prevalence : 0.75
## Balanced Accuracy : 1.00
##
## 'Positive' Class : Good
##
# Confidence Interval is around 97 %
# no of nodes of trees
hist(treesize(rf1), main = "No of Nodes for the trees", col = "green")
# Majority of the trees has an average number of more than 80 nodes.
# Variable Importance :
varImpPlot(rf1)
# Mean Decrease Accuracy graph shows that how worst the model performs without each variable.
# say Taxable.Income is the most important variable for prediction.on looking at City population,it has no value.
# MeanDecrease gini graph shows how much by average the gini decreases if one of those nodes were
# removed. Taxable.Income is very important and Urban is not that important.
varImpPlot(rf1 ,Sort = T, n.var = 5, main = "Top 5 -Variable Importance")
## Warning in mtext(labs, side = 2, line = loffset, at = y, adj = 0, col =
## color, : "Sort" is not a graphical parameter
## Warning in title(main = main, xlab = xlab, ylab = ylab, ...): "Sort" is not
## a graphical parameter
## Warning in mtext(labs, side = 2, line = loffset, at = y, adj = 0, col =
## color, : "Sort" is not a graphical parameter
## Warning in title(main = main, xlab = xlab, ylab = ylab, ...): "Sort" is not
## a graphical parameter
# Quantitative values
importance(rf1)
## Good Risky MeanDecreaseAccuracy
## Undergrad 0.9697747 0.7452060 1.0746869
## Marital.Status 2.3190683 -0.1647737 1.6152058
## Taxable.Income 57.7118526 74.8748950 64.0732787
## City.Population 0.6655007 -2.5773995 -1.4062559
## Work.Experience 0.1951129 -1.0500176 -0.8912347
## Urban 1.2305502 0.1132935 0.9469538
## MeanDecreaseGini
## Undergrad 0.6916259
## Marital.Status 1.8147131
## Taxable.Income 115.2485651
## City.Population 4.5044196
## Work.Experience 3.0332879
## Urban 0.5027630
varUsed(rf) # which predictor variables are actually used in the random forest.
## [1] 315 492 1719 1275 962 325
# Partial Dependence Plot
partialPlot(rf1, train, Taxable.Income, "Good")
# On that graph, i see that if the taxable Income is 30000 or greater,
# than they are good customers else those are risky customers.
# Extract single tree from the forest :
tr1 <- getTree(rf1, 2, labelVar = TRUE)
# Multi Dimension scaling plot of proximity Matrix
MDSplot(rf1, FC$Risky_Good)
## Warning in RColorBrewer::brewer.pal(nlevs, "Set1"): minimal value for n is 3, returning requested palette with 3 different levels
