Introduction

This data set is about a direct marketing case from the insurance sector which was to predict policy ownership. It is about predicting who would be interested in buying a caravan insurance policy. This data set was used in the second edition of the Computational intelligence and Learning(CoIL) competition Challenge in the Year 2000, organized by CoIL cluster, which is a cooperation between four EU funded Networks of Excellence which represent the areas of neural networks (NeuroNet), fuzzy systems (ERUDIT), evolutionary computing (EvoNet) and machine learning (MLNet) and it is owned and donated by Peter van der Putten of the Dutch data mining company Sentient Machine Research, Baarsjesweg 224 1058 AA Amsterdam The Netherlands +31 20 6186927 putten@liacs.nl and is based on real world business problem. TIC (The Insurance Company) Benchmark Homepage (http://www.liacs.nl/~putten/library/cc2000) was donated on March 7, 2000.

Relevant Papers

P. van der Putten and M. van Someren (eds). CoIL Challenge 2000: The Insurance Company Case. Published by Sentient Machine Research, Amsterdam. Also a Leiden Institute of Advanced Computer Science Technical Report 2000-09. June 22, 2000.

SUMMARY ABOUT DATASET

NO OF OBSERVATIONS: 5822 real customer records

NO OF VARIABLES: 86 Nos.

Each real customer record consists of 86 variables, containing sociodemographic data (variables 1-43) and product ownership data (variables 44-86). The sociodemographic data is derived from zip codes. All customers living in areas with the same zip code have the same sociodemographic attributes. Variable 86 (Purchase), “CARAVAN: Number of mobile home policies”, is the target variable which indicates whether the customer purchase a caravan insurance policy or not.

TASK

Predict which customers are potentially interested in a caravan insurance policy (Prediction or Classification).

PREDICTION TASK

To predict whether a customer is interested in a caravan insurance policy from other data about the customer. Information about customers consists of 86 variables and includes product usage data and socio-demographic data derived from zip area codes. The data was supplied by the Dutch data mining company Sentient Machine Research and is based on a real world business problem. The training set contains over 5000 descriptions of customers, including the information of whether or not they have a caravan insurance policy. A test set contains 4000 customers. In the prediction task, the underlying problem is to the find the subset of customers with a probability of having a caravan insurance policy above some boundary probability. The known policyholders can then be removed and the rest receives a mailing. The boundary depends on the costs and benefits such as of the costs of mailing and benefit of selling insurance policies. To approximate this problem, we want to find the set of 800 customers in the test set of 4000 customers that contains the most caravan policy owners. For each solution submitted, the number of actual policyholders will be counted and this gives the score of a solution.

library(ISLR)

## PIE CHART OF YES/NO FOR PURCHASE OF CARAVAN POLICY BY CUSTOMERS

r a<-table(Caravan$Purchase) a

## ## No Yes ## 5474 348

colors=c("red","green")
col=colors
pie(a,main = "CUSTOMERS OF CARAVAN POLICY",col=colors)
box()

OBSERVATION FOR PIE CHART The above piechart shows the number of customers who purchased(Yes) the Caravan policy which is 348 and who have not purchased(NO) the Caravan policy which is 5474

# BAR AND PIE CHARTS SHOWING CORRELATION OF CUSTOMERS WHO PURCHASED CARAVAN POLICY AND VARIOUS VARIABLES

## BAR CHARTS AND PIE CHARTS SHOWING PURCHASE OF CARAVAN POLICY BY CUSTOMERS AGAINST PRODUCT USAGE(POLICY OWNERSHIP) DATA VARIABLES

### 1.VARIABLE - NUMBER OF BOAT POLICIES

a<-table(Caravan$APLEZIER[Caravan$Purchase=="Yes"])
a
## 
##   0   1   2 
## 335  12   1
barplot(a,border="dark blue",main = "PURCHASE OF CARAVAN POLICY vs NUMBER OF BOAT POLICIES",xlab = "Number of boat policies",ylab = "Number of customers")

OBSERVATION OF CUSTOMER TYPE

In the above barplot, We come to know that the Customers who have not purchased the boat policy(0) are more likely to purchase the Caravan policy

### 2. VARIABLE - NUMBER OF SOCIAL SECURITY INSURANCE POLICIES

a<-table(Caravan$ABYSTAND[Caravan$Purchase=="Yes"])
a
## 
##   0   1 
## 332  16
barplot(a,border="dark blue",main = "PURCHASE OF CARAVAN POLICY vs NO. OF SS INSURANCE POLICIES",xlab = "Number of social security insurance policies",ylab = "Number of customers")

OBSERVATION OF CUSTOMER TYPE

In the above barplot, We come to know that the Customers who have not purchased social security insurance policy(0) are more likely to purchase the Caravan policy

### 3. VARIABLE - CONTRIBUTION CAR POLICIES

a<-table(Caravan$PPERSAUT[Caravan$Purchase=="Yes"])
a
## 
##   0   5   6 
##  72  14 262
colors=c("blue","red","green")
col=colors
pie(a,main ="PURCHASE OF CARAVAN POLICY vs CONTRIBUTION CAR POLICIES",col=colors)
box()

OBSERVATION OF CUSTOMER TYPE

In the above piechart, We come to know that the Customers who pay car policy premium averagely from $1000 to $4999(6)  are more likely to purchase the Caravan policy

### 4. VARIABLE - Number of fire policies

a<-table(Caravan$ABRAND[Caravan$Purchase=="Yes"])
a
## 
##   0   1   2 
## 109 232   7
colors=c("orange","violet","yellow")
col=colors
pie(a,main ="PURCHASE OF CARAVAN POLICY vs NUMBER OF FIRE POLICIES",col=colors)
box()

OBSERVATION OF CUSTOMER TYPE

In the above piechart, We come to know that the Customers who purchase only one fire policy are more likely to purchase the Caravan policy

## CHARTS SHOWING PURCHASE OF CARAVAN POLICY BY CUSTOMERS AGAINST SOCIODEMOGRAPHIC DATA VARIABLES

### 1. VARIABLE - CUSTOMER SUBTYPE

r a<-table(Caravan$MOSTYPE[Caravan$Purchase=="Yes"]) a

## ## 1 2 3 4 5 6 7 8 9 10 11 12 13 20 22 23 24 25 26 27 29 30 31 32 33 ## 13 6 25 2 2 12 3 51 12 9 9 16 13 2 4 4 5 2 1 1 2 4 6 8 46 ## 34 35 36 37 38 39 41 ## 9 8 16 10 23 19 5

r barplot(a,border="dark blue",main = "PURCHASE OF CARAVAN POLICY vs CUSTOMER SUBTYPE",xlab="Customer subtype",ylab="Number of customers")

OBSERVATION OF CUSTOMER TYPE

In the above barplot, Customers of various subtype of about 41 labels are taken. Customers belonging to subtype 8(Middle class families) & subtype 33(lower class with large families) are more likely to purchase the Caravan policy

### 2. VARIABLE - AVG AGE (Age group)

r a<-table(Caravan$MGEMLEEF[Caravan$Purchase=="Yes"]) a

## ## 1 2 3 4 5 6 ## 1 87 183 64 12 1

names(a)=c("20 to 30","30 to 40","40 to 50","50 to 60","60 to 70","70 to 80")
barplot(a,col=rainbow(6),main = "PURCHASE OF CARAVAN POLICY vs AVE AGE",xlab="Avg age or Age group",ylab="Number of customers")

OBSERVATION FOR AVG AGE

In the above barplot, customers of various age group is taken and it is plotted against the customers who have said yes to buy caravan policy. The customers belonging to age group of 40-50 are more likely to purchase the caravan policy

### 3. VARIABLE - PURCHASING POWER CLASS

r a<-table(Caravan$MKOOPKLA[Caravan$Purchase=="Yes"]) a

## ## 1 2 3 4 5 6 7 8 ## 18 15 71 46 30 66 67 35

barplot(a,col=rainbow(7),main = "PURCHASE OF CARAVAN POLICY vs PURCHASING POWER CLASS",xlab = "Purchasing power class",ylab = "Number of customers")

OBSERVATION OF CUSTOMER TYPE

In the above barplot, We come to know that the Customers who are of High status seniors(3) are more likely to purchase the Caravan policy with Dinki's (double income no kids)(7) people and Career and childcare class(6) coming a close second and third respectively

### 4. VARIABLE - AVERAGE INCOME

a<-table(Caravan$MINKGEM[Caravan$Purchase=="Yes"])
a
## 
##   1   2   3   4   5   6   7   8 
##   1  20  69 139  70  24  17   8
pie(a,col=rainbow(7),main ="PURCHASE OF CARAVAN POLICY vs AVERAGE INCOME")
box()

OBSERVATION OF CUSTOMER TYPE

In the above piechart, We come to know that the middle income Customers who are of the average income between $200 to $499(4) are more likely and the Customers who are of the average income between $100 to $199(3) and between $500 to $999(5) are likely to purchase the Caravan policy

### 5. VARIABLE - CUSTOMER MAIN TYPE

b<-table(Caravan$MOSHOOFD[Caravan$Purchase=="Yes"])
b
## 
##  1  2  3  5  6  7  8  9 10 
## 48 66 59 15  4 20 89 42  5
colors=c("violet","yellow","blue","red","brown","orange","green")
color=colors
pie(b,col=colors,main ="PURCHASE OF CARAVAN POLICY vs CUSTOMER MAIN TYPE")
box()

OBSERVATION OF CUSTOMER TYPE

In the above Pie chart, Customers of various maintype of about 10 labels are taken. Customers belonging to maintype 8(Family with grown ups) & maintype 2(Driven Growers) are more likely to purchase the Caravan policy

PREDICTION MODELS USING ALGORITHMS

RPART GLM C50 - Rules and Trees ZERO R ————————————————————————————————

MODEL No 1

#PREDICTION USING ALGORITHM( RPART)

library(rpart)
library(rattle)
## Loading required package: RGtk2
## Rattle: A free graphical interface for data mining with R.
## Version 3.5.0 Copyright (c) 2006-2015 Togaware Pty Ltd.
## Type 'rattle()' to shake, rattle, and roll your data.
library(rpart.plot)
library(RColorBrewer)
library(crossval)
library(gplots)
## 
## Attaching package: 'gplots'
## 
## The following object is masked from 'package:stats':
## 
##     lowess
library(vcd)
## Loading required package: grid
## 
## Attaching package: 'vcd'
## 
## The following object is masked from 'package:ISLR':
## 
##     Hitters
library(Metrics)

######## Read the Caravan.csvfile

d1<- read.csv("C:/Users/S.RAJKUMAR/Desktop/Caravan2.csv")
d1.ori<-d1
set.seed(99)

tr <- d1.ori[sample(row.names(d1.ori), size = round(nrow(d1.ori)*0.5)), ]
te <- d1.ori[!(row.names(d1.ori) %in% row.names(tr)), ]

Reset the original training and test data - just to be sure

tr1 <- tr
te1  <- te
te2 <-te

zero r startgey no one will purchase

te2$Purchase <- rep(0,nrow(te2))

buliding the tree

tr1$Purchase = as.factor(tr1$Purchase)
fit1 <- rpart(formula=Purchase ~ .,data=tr1,control=rpart.control(minsplit=20, minbucket=1, cp=0.008))
fit1
## n= 2911 
## 
## node), split, n, loss, yval, (yprob)
##       * denotes terminal node
## 
##  1) root 2911 181 0 (0.93782205 0.06217795)  
##    2) PPERSAUT< 5.5 1755  53 0 (0.96980057 0.03019943) *
##    3) PPERSAUT>=5.5 1156 128 0 (0.88927336 0.11072664)  
##      6) MOSTYPE>=12.5 785  56 0 (0.92866242 0.07133758) *
##      7) MOSTYPE< 12.5 371  72 0 (0.80592992 0.19407008)  
##       14) PBRAND< 3.5 213  24 0 (0.88732394 0.11267606)  
##         28) MBERHOOG< 5.5 181  15 0 (0.91712707 0.08287293) *
##         29) MBERHOOG>=5.5 32   9 0 (0.71875000 0.28125000)  
##           58) MBERMIDD< 1.5 23   3 0 (0.86956522 0.13043478) *
##           59) MBERMIDD>=1.5 9   3 1 (0.33333333 0.66666667) *
##       15) PBRAND>=3.5 158  48 0 (0.69620253 0.30379747)  
##         30) MBERMIDD< 6.5 142  37 0 (0.73943662 0.26056338) *
##         31) MBERMIDD>=6.5 16   5 1 (0.31250000 0.68750000) *
gc()
##           used (Mb) gc trigger (Mb) max used (Mb)
## Ncells  617193 33.0    1168576 62.5   774200 41.4
## Vcells 1850429 14.2    3251679 24.9  3237369 24.7
fancyRpartPlot(fit1)

NAMES INFORMATION VALUES 1 labels
PPERSAUT car policy 1-8 values -
MOSTYPE Customer subtype 1-41 FYE,12(affluent young)
PBRAND fire policy ** (0-7) values** -
MBERHOOG High status (0-9)values -
MBERMIDD Middle management (0-9)values - 
printcp(fit1)
## 
## Classification tree:
## rpart(formula = Purchase ~ ., data = tr1, control = rpart.control(minsplit = 20, 
##     minbucket = 1, cp = 0.008))
## 
## Variables actually used in tree construction:
## [1] MBERHOOG MBERMIDD MOSTYPE  PBRAND   PPERSAUT
## 
## Root node error: 181/2911 = 0.062178
## 
## n= 2911 
## 
##          CP nsplit rel error xerror     xstd
## 1 0.0082873      0   1.00000  1.000 0.071982
## 2 0.0080000      6   0.95028  1.105 0.075402
print(fit1)
## n= 2911 
## 
## node), split, n, loss, yval, (yprob)
##       * denotes terminal node
## 
##  1) root 2911 181 0 (0.93782205 0.06217795)  
##    2) PPERSAUT< 5.5 1755  53 0 (0.96980057 0.03019943) *
##    3) PPERSAUT>=5.5 1156 128 0 (0.88927336 0.11072664)  
##      6) MOSTYPE>=12.5 785  56 0 (0.92866242 0.07133758) *
##      7) MOSTYPE< 12.5 371  72 0 (0.80592992 0.19407008)  
##       14) PBRAND< 3.5 213  24 0 (0.88732394 0.11267606)  
##         28) MBERHOOG< 5.5 181  15 0 (0.91712707 0.08287293) *
##         29) MBERHOOG>=5.5 32   9 0 (0.71875000 0.28125000)  
##           58) MBERMIDD< 1.5 23   3 0 (0.86956522 0.13043478) *
##           59) MBERMIDD>=1.5 9   3 1 (0.33333333 0.66666667) *
##       15) PBRAND>=3.5 158  48 0 (0.69620253 0.30379747)  
##         30) MBERMIDD< 6.5 142  37 0 (0.73943662 0.26056338) *
##         31) MBERMIDD>=6.5 16   5 1 (0.31250000 0.68750000) *
plot(fit1)
text(fit1)

fit1$cptable[which.min(fit1$cptable[,"xerror"]),"CP"]
## [1] 0.008287293
Prediction<-predict(fit1,te1,type="class")

Compare with base model

Update the prediction

te2$Purchase <- Prediction

Pred = factor(as.factor(te2$Purchase), c(0, 1), labels = c("Not purchased", "Purchased"))
Actual = factor(as.factor(te1$Purchase), c(0, 1), labels = c("Not purchased", "Purchased"))
table(te1$Purchase)                     
## 
##    0    1 
## 2744  167
cm1 = confusionMatrix(Actual,Pred, negative = "Not purchased")
cm1
##   FP   TP   TN   FN 
##   23    6 2721  161 
## attr(,"negative")
## [1] "Not purchased"
# corresponding accuracy, sensitivity etc.
diagnosticErrors(cm1)
##        acc       sens       spec        ppv        npv        lor 
## 0.93679148 0.03592814 0.99161808 0.20689655 0.94413602 1.48361563 
## attr(,"negative")
## [1] "Not purchased"
# Compute the classification error
ce(Actual,Pred)                     
## [1] 0.06320852
Model No 2

#PREDICTION USING ALGORITHM(GLM)

Variable Used

NAMES INFORMATION VALUES 1 labels
MOSHOOFD customer main typr 1-10 values -
MSKB1 Social Class B1 0-41 -
PWAPART Pvt 3rd Party Ins (0-9) values - 
library(ggplot2)
library(MASS)
library(splines)
library(mgcv)
## Loading required package: nlme
## This is mgcv 1.8-7. For overview type 'help("mgcv-package")'.
library(crossval)

Caravan2<- read.csv("C:/Users/S.RAJKUMAR/Desktop/Caravan2.csv")
Caravan.ori <- Caravan2

set.seed(11)
train <- Caravan.ori[sample(row.names(Caravan.ori), size = round(nrow(Caravan.ori)*0.7)), ]
test <- Caravan.ori[!(row.names(Caravan.ori) %in% row.names(train)), ]

train.ori <-train
test.ori<-test

train2<-train
test2<-test

##No one purchased

test2$Purchase <- rep(0, nrow(test2))

glm.logistic <- glm(Purchase ~ MOSHOOFD + MSKB1+PWAPART,
                    family = "binomial", data = train)
Prediction.prob <- predict(glm.logistic, newdata = test, type="response")
Prediction <- round(Prediction.prob,0)

####

test2$Purchase <- Prediction

Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

cm5 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm5
##   FP   TP   TN   FN 
##    0    0 1638  109 
## attr(,"negative")
## [1] "Not Purchased"
diagnosticErrors(cm5)
##       acc      sens      spec       ppv       npv       lor 
## 0.9376073 0.0000000 1.0000000       NaN 0.9376073       NaN 
## attr(,"negative")
## [1] "Not Purchased"
ce(Actual.Outcome,Our.Prediction)
## [1] 0.06239267
glm.fit <- glm(Purchase ~ MOSHOOFD + PWAPART + MSKB1,
               family = "binomial", data = train)

inv.logit <- function(x) exp(x) / (1 + exp(x))

glm.pred <- predict(glm.fit, newdata = test, se.fit = TRUE)

pred <- data.frame(mean = inv.logit(glm.pred$fit),
                   lo = inv.logit(glm.pred$fit - 2 * glm.pred$se.fit),
                   hi = inv.logit(glm.pred$fit + 2 * glm.pred$se.fit),
                   Purchase = test$Purchase)
pred <- pred[order(pred$mean), ]
pred$id <- seq_along(pred$mean)
row.names(pred) <- NULL

sim.data <- expand.grid(MSKB1 = 2, MOSHOOFD = 8,
                        PWAPART = 0)

pred <- predict(glm.fit, newdata = sim.data, se.fit = TRUE)
sim.data$mean <- inv.logit(pred$fit)
sim.data$lo <- inv.logit(pred$fit - 2 * pred$se.fit)
sim.data$hi <- inv.logit(pred$fit + 2 * pred$se.fit)
Prediction <- round(Prediction.prob,0)

Compare with base model

# Update the prediction with out model output

test2$Purchase <- Prediction

Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm6 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm6
##   FP   TP   TN   FN 
##    0    0 1638  109 
## attr(,"negative")
## [1] "Not Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm6)
##       acc      sens      spec       ppv       npv       lor 
## 0.9376073 0.0000000 1.0000000       NaN 0.9376073       NaN 
## attr(,"negative")
## [1] "Not Purchased"
Model No 3

#PREDICTION USING ALGORITHM(C50) Variables Used

NAMES INFORMATION VALUES labels
PPLEZIER Cont to Boat Policy 0-6 values -
PBYSTAND Cont to Social Sec 0-5 -
APLEZIER No of boat policies 0-3 -
ABYSTAND No of Social Sec (0-2) values -

# Strategy 7 - C50 trees (rules) basic

# Read the Caravan data from Caravan2.csv

Caravan.ori <-Caravan2

set.seed(11)
train <- Caravan.ori[sample(row.names(Caravan.ori), size = round(nrow(Caravan.ori)*0.7)), ]
test <- Caravan.ori[!(row.names(Caravan.ori) %in% row.names(train)), ]

# Creating backup of test and train data for later use. Not modifying .ori files as a rule

train.ori <-train
test.ori<-test
train2<-train
test2<-test
library(crossval)
library(gplots)
library(vcd)
library(Metrics)
library(C50)

# Resetting the original training and test data - just to be sure

train <- train.ori
test  <- test.ori
test2 <-test

# Also resetting the test2 data with no one purchased ZeroR strategy

test2$Purchase <- rep(0, nrow(test2))

combinedData1 <- Caravan.ori[,-7]
combinedData2 <- combinedData1[,-6]
combinedData <- combinedData2[,-5]

combinedData$Purchase <- factor(combinedData$Purchase)


set.seed(11)
train <- combinedData[sample(row.names(combinedData), size = round(nrow(combinedData)*0.7)), ]
test <- combinedData[!(row.names(combinedData) %in% row.names(train)), ]
table(train$Purchase)
## 
##    0    1 
## 3836  239
table(test$Purchase)
## 
##    0    1 
## 1638  109
C50.Rules <- C5.0(Purchase~PPLEZIER+PBYSTAND+APLEZIER+ABYSTAND, data=train, rules = FALSE)

Prediction <- predict(C50.Rules,test)

################################## Comparing with base model # Updating the prediction with out model output

test2$Purchase <- Prediction

Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm7 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm7
##   FP   TP   TN   FN 
##    0    0 1638  109 
## attr(,"negative")
## [1] "Not Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm7)
##       acc      sens      spec       ppv       npv       lor 
## 0.9376073 0.0000000 1.0000000       NaN 0.9376073       NaN 
## attr(,"negative")
## [1] "Not Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.06239267

# Strategy 8 - Tree model of C50

C50.Tree <- C5.0(train[,-86],train$Purchase)

Prediction <- predict(C50.Tree,test)

################################## Comparing with base model # Updating the prediction with out model output

test2$Purchase <- Prediction

Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm8 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm8
##   FP   TP   TN   FN 
##    0  109 1638    0 
## attr(,"negative")
## [1] "Not Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm8)
##  acc sens spec  ppv  npv  lor 
##    1    1    1    1    1  Inf 
## attr(,"negative")
## [1] "Not Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0

# Strategy 9 - Tree model of C50 with min of 2 items in tree edges 

C50.Tree.small <- C5.0(train[,-3],train$Purchase,
control = C5.0Control(minCases = 2))

Prediction <- predict(C50.Tree.small,test)

################################## Comparing with base model # Updating the prediction with out model output

test2$Purchase <- Prediction

Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm9 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm9
##   FP   TP   TN   FN 
##    0  109 1638    0 
## attr(,"negative")
## [1] "Not Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm9)
##  acc sens spec  ppv  npv  lor 
##    1    1    1    1    1  Inf 
## attr(,"negative")
## [1] "Not Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0

# Strategy 10 - Tree model of C50 Rule & Tree Model 

library(ISLR)
library(C50)
treeModel <- C5.0(x = Caravan[, -86], y = Caravan$Purchase)
treeModel
## 
## Call:
## C5.0.default(x = Caravan[, -86], y = Caravan$Purchase)
## 
## Classification Tree
## Number of samples: 5822 
## Number of predictors: 85 
## 
## Tree size: 1 
## 
## Non-standard options: attempt to group attributes
summary(treeModel)
## 
## Call:
## C5.0.default(x = Caravan[, -86], y = Caravan$Purchase)
## 
## 
## C5.0 [Release 2.07 GPL Edition]      Sat Oct 17 21:55:23 2015
## -------------------------------
## 
## Class specified by attribute `outcome'
## 
## Read 5822 cases (86 attributes) from undefined.data
## 
## Decision tree:
##  No (5822/348)
## 
## 
## Evaluation on training data (5822 cases):
## 
##      Decision Tree   
##    ----------------  
##    Size      Errors  
## 
##       1  348( 6.0%)   <<
## 
## 
##     (a)   (b)    <-classified as
##    ----  ----
##    5474          (a): class No
##     348          (b): class Yes
## 
## 
## Time: 0.2 secs
ruleModel <- C5.0(Purchase ~ ., data = Caravan, rules = FALSE)
ruleModel
## 
## Call:
## C5.0.formula(formula = Purchase ~ ., data = Caravan, rules = FALSE)
## 
## Classification Tree
## Number of samples: 5822 
## Number of predictors: 85 
## 
## Tree size: 1 
## 
## Non-standard options: attempt to group attributes
summary(ruleModel)
## 
## Call:
## C5.0.formula(formula = Purchase ~ ., data = Caravan, rules = FALSE)
## 
## 
## C5.0 [Release 2.07 GPL Edition]      Sat Oct 17 21:55:24 2015
## -------------------------------
## 
## Class specified by attribute `outcome'
## 
## Read 5822 cases (86 attributes) from undefined.data
## 
## Decision tree:
##  No (5822/348)
## 
## 
## Evaluation on training data (5822 cases):
## 
##      Decision Tree   
##    ----------------  
##    Size      Errors  
## 
##       1  348( 6.0%)   <<
## 
## 
##     (a)   (b)    <-classified as
##    ----  ----
##    5474          (a): class No
##     348          (b): class Yes
## 
## 
## Time: 0.2 secs
treeModel <- C5.0(x = Caravan[, -86], y = Caravan$Purchase,
                  control = C5.0Control(winnow = FALSE))
summary(treeModel)
## 
## Call:
## C5.0.default(x = Caravan[, -86], y = Caravan$Purchase, control
##  = C5.0Control(winnow = FALSE))
## 
## 
## C5.0 [Release 2.07 GPL Edition]      Sat Oct 17 21:55:25 2015
## -------------------------------
## 
## Class specified by attribute `outcome'
## 
## Read 5822 cases (86 attributes) from undefined.data
## 
## Decision tree:
##  No (5822/348)
## 
## 
## Evaluation on training data (5822 cases):
## 
##      Decision Tree   
##    ----------------  
##    Size      Errors  
## 
##       1  348( 6.0%)   <<
## 
## 
##     (a)   (b)    <-classified as
##    ----  ----
##    5474          (a): class No
##     348          (b): class Yes
## 
## 
## Time: 0.2 secs
treeModel <- C5.0(x = Caravan[, -86], y = Caravan$Purchase,
                  control = C5.0Control(winnow = FALSE, minCases = 5))
summary(treeModel)
## 
## Call:
## C5.0.default(x = Caravan[, -86], y = Caravan$Purchase, control
##  = C5.0Control(winnow = FALSE, minCases = 5))
## 
## 
## C5.0 [Release 2.07 GPL Edition]      Sat Oct 17 21:55:26 2015
## -------------------------------
## 
## Class specified by attribute `outcome'
## 
## Read 5822 cases (86 attributes) from undefined.data
## 
## Decision tree:
##  No (5822/348)
## 
## 
## Evaluation on training data (5822 cases):
## 
##      Decision Tree   
##    ----------------  
##    Size      Errors  
## 
##       1  348( 6.0%)   <<
## 
## 
##     (a)   (b)    <-classified as
##    ----  ----
##    5474          (a): class No
##     348          (b): class Yes
## 
## 
## Time: 0.2 secs
Model No 4

#PREDICTION USING ALGORITHM(ZERO R)

Variables Used

NAMES INFORMATION VALUES labels
All Variable All Variables -
MGEMLEEF Avg age of customer 1-6 values 3 - 40 to 50 Yrs
MOSTYPE Customer Sub Type 0-41 -33 with more records
PBRAND Cont Fire Plocy 0-8 -
PPERSAUT No of Car Policy (0, 4-8) - 
library(crossval)
library(gplots)
library(vcd)
library(Metrics) 
Caravan2 <- read.csv("C:/Users/S.RAJKUMAR/Desktop/Caravan2.csv")

# Read the Caravan data from Caravan2.csv

Caravan.ori <- Caravan2
set.seed(11)
train <- Caravan.ori[sample(row.names(Caravan.ori), size = round(nrow(Caravan.ori)*0.5)), ]
test <- Caravan.ori[!(row.names(Caravan.ori) %in% row.names(train)), ]

#Create backup of test and train data for later use. Do not modify .ori files as a rule

train.ori <-train
test.ori<-test

train2<-train
test2<-test

# Looking at NO. of people who Purchased or not the Caravan policy

table(train$Purchase)
## 
##    0    1 
## 2745  166
table(test$Purchase)
## 
##    0    1 
## 2729  182
prop.table(table(train$Purchase))
## 
##          0          1 
## 0.94297492 0.05702508

# Strategy 11 - ZeroR model # Using ZeroR algorithm and solving it. # Creating new column in test set with our prediction every one has purchased

test2$Purchase <- rep(1, nrow(test2))

Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm11 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Purchased")
cm11
##   FP   TP   TN   FN 
##    0    0  182 2729 
## attr(,"negative")
## [1] "Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm11)
##        acc       sens       spec        ppv        npv        lor 
## 0.06252147 0.00000000 1.00000000        NaN 0.06252147        NaN 
## attr(,"negative")
## [1] "Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.9374785

#Strategy 12 - ZeroR model

# Creating new column in test set with our prediction no one purchased

test2$Purchase <- rep(0, nrow(test2))

Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm12 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm12
##   FP   TP   TN   FN 
##    0    0 2729  182 
## attr(,"negative")
## [1] "Not Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm12)
##       acc      sens      spec       ppv       npv       lor 
## 0.9374785 0.0000000 1.0000000       NaN 0.9374785       NaN 
## attr(,"negative")
## [1] "Not Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.06252147

# Strategy 13 - Customer Sub Type

# Resetting the original training and test data - just to be sure

train <- train.ori
test  <- test.ori
test2 <-test

# Also resetting the test2 data with no one purchased ZeroR strategy

test2$Purchase <- rep(0, nrow(test2))

summary(train$MOSTYPE)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##    1.00   10.00   30.00   24.23   35.00   41.00
prop.table(table(train$MOSTYPE, train$Purchase))
##     
##                 0            1
##   1  0.0178632772 0.0013740982
##   2  0.0151150807 0.0013740982
##   3  0.0415664720 0.0027481965
##   4  0.0106492614 0.0003435246
##   5  0.0082445895 0.0003435246
##   6  0.0195809000 0.0017176228
##   7  0.0089316386 0.0010305737
##   8  0.0467193404 0.0068704912
##   9  0.0446581931 0.0024046719
##   10 0.0254208176 0.0017176228
##   11 0.0237031948 0.0013740982
##   12 0.0168327035 0.0027481965
##   13 0.0271384404 0.0020611474
##   15 0.0006870491 0.0000000000
##   16 0.0034352456 0.0000000000
##   17 0.0013740982 0.0000000000
##   18 0.0034352456 0.0000000000
##   19 0.0006870491 0.0000000000
##   20 0.0037787702 0.0000000000
##   21 0.0020611474 0.0000000000
##   22 0.0154586053 0.0003435246
##   23 0.0425970457 0.0003435246
##   24 0.0291995878 0.0006870491
##   25 0.0147715562 0.0003435246
##   26 0.0089316386 0.0003435246
##   27 0.0075575404 0.0000000000
##   28 0.0048093439 0.0000000000
##   29 0.0147715562 0.0003435246
##   30 0.0202679492 0.0006870491
##   31 0.0357265544 0.0013740982
##   32 0.0223290965 0.0006870491
##   33 0.1308828581 0.0089316386
##   34 0.0305736860 0.0013740982
##   35 0.0350395053 0.0013740982
##   36 0.0336654071 0.0034352456
##   37 0.0216420474 0.0013740982
##   38 0.0546204054 0.0051528684
##   39 0.0570250773 0.0027481965
##   40 0.0109927860 0.0000000000
##   41 0.0302301615 0.0013740982
prop.table(table(train$MOSTYPE, train$Purchase), 1)
##     
##               0          1
##   1  0.92857143 0.07142857
##   2  0.91666667 0.08333333
##   3  0.93798450 0.06201550
##   4  0.96875000 0.03125000
##   5  0.96000000 0.04000000
##   6  0.91935484 0.08064516
##   7  0.89655172 0.10344828
##   8  0.87179487 0.12820513
##   9  0.94890511 0.05109489
##   10 0.93670886 0.06329114
##   11 0.94520548 0.05479452
##   12 0.85964912 0.14035088
##   13 0.92941176 0.07058824
##   15 1.00000000 0.00000000
##   16 1.00000000 0.00000000
##   17 1.00000000 0.00000000
##   18 1.00000000 0.00000000
##   19 1.00000000 0.00000000
##   20 1.00000000 0.00000000
##   21 1.00000000 0.00000000
##   22 0.97826087 0.02173913
##   23 0.99200000 0.00800000
##   24 0.97701149 0.02298851
##   25 0.97727273 0.02272727
##   26 0.96296296 0.03703704
##   27 1.00000000 0.00000000
##   28 1.00000000 0.00000000
##   29 0.97727273 0.02272727
##   30 0.96721311 0.03278689
##   31 0.96296296 0.03703704
##   32 0.97014925 0.02985075
##   33 0.93611794 0.06388206
##   34 0.95698925 0.04301075
##   35 0.96226415 0.03773585
##   36 0.90740741 0.09259259
##   37 0.94029851 0.05970149
##   38 0.91379310 0.08620690
##   39 0.95402299 0.04597701
##   40 1.00000000 0.00000000
##   41 0.95652174 0.04347826

################################## Comparing with base model

# Updating the prediction to say that Subtype will Purchase

test2$Purchase[test2$MGEMLEEF ==3] <- 1
Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm13 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Purchased")
cm13
##   FP   TP   TN   FN 
##   90 1294   92 1435 
## attr(,"negative")
## [1] "Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm13)
##         acc        sens        spec         ppv         npv         lor 
##  0.47612504  0.47416636  0.50549451  0.93497110  0.06024885 -0.08144775 
## attr(,"negative")
## [1] "Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.523875

########################### Strategy 14 - Customer Sub Type

#################### Resetting the original training and test data - just to be sure

train <- train.ori
test  <- test.ori
test2 <-test

# Also resetting the test2 data with no one purchased ZeroR strategy

test2$Purchase <- rep(0, nrow(test2))
summary(train$MOSTYPE)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##    1.00   10.00   30.00   24.23   35.00   41.00

################################## Comparing with base model

################# Updating the prediction to say that Subtype will not Purchase

test2$Purchase[test2$MGEMLEEF==3] <- 0
Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm14 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm14
##   FP   TP   TN   FN 
##    0    0 2729  182 
## attr(,"negative")
## [1] "Not Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm14)
##       acc      sens      spec       ppv       npv       lor 
## 0.9374785 0.0000000 1.0000000       NaN 0.9374785       NaN 
## attr(,"negative")
## [1] "Not Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.06252147

################### Strategy 15 - Customer Sub Type 33

#################### Resetting the original training and test data - just to be sure

train <- train.ori
test  <- test.ori
test2 <-test

# Also resetting the test2 data with no one purchased ZeroR strategy

test2$Purchase <- rep(0, nrow(test2))
summary(train$MOSTYPE)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##    1.00   10.00   30.00   24.23   35.00   41.00

################################## Comparing with base model

# Updating the prediction to say that Subtype 33 will Purchase

test2$Purchase[test2$MOSTYPE==33] <- 1
Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm15 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Purchased")
cm15
##   FP   TP   TN   FN 
##  162 2346   20  383 
## attr(,"negative")
## [1] "Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm15)
##         acc        sens        spec         ppv         npv         lor 
##  0.81277911  0.85965555  0.10989011  0.93540670  0.04962779 -0.27943202 
## attr(,"negative")
## [1] "Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.1872209

# Strategy 16 - Customer Sub Type 33

# Resetting the original training and test data - just to be sure

train <- train.ori
test  <- test.ori
test2 <-test

# Also resetting the test2 data with no one purchased ZeroR strategy

test2$Purchase <- rep(0, nrow(test2))

summary(train$MOSTYPE)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##    1.00   10.00   30.00   24.23   35.00   41.00

################################## Comparing with base model

# Updating the prediction to say that Subtype 33 will not Purchase

test2$Purchase[test2$MOSTYPE==33] <- 0


Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm16 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm16
##   FP   TP   TN   FN 
##    0    0 2729  182 
## attr(,"negative")
## [1] "Not Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm16)
##       acc      sens      spec       ppv       npv       lor 
## 0.9374785 0.0000000 1.0000000       NaN 0.9374785       NaN 
## attr(,"negative")
## [1] "Not Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.06252147

# Strategy 17 - Contribution to Fire Policy

# Resetting the original training and test data - just to be sure

train <- train.ori
test  <- test.ori
test2 <-test

# Also reset the test2 data with no one purchased ZeroR strategy

test2$Purchase <- rep(0, nrow(test2))

################################## Comparing with base model

# Updating the prediction to say that Customer will Purchase

test2$Purchase[test2$PBRAND] <- 1


Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm17 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Purchased")
cm17
##   FP   TP   TN   FN 
##  182 2721    0    8 
## attr(,"negative")
## [1] "Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm17)
##       acc      sens      spec       ppv       npv       lor 
## 0.9347303 0.9970685 0.0000000 0.9373062 0.0000000      -Inf 
## attr(,"negative")
## [1] "Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.06526967

# Strategy 18 - Contribution to Fire Policy

# Resetting the original training and test data - just to be sure

train <- train.ori
test  <- test.ori
test2 <-test

# Also resetting the test2 data with no one purchased ZeroR strategy

test2$Purchase <- rep(0, nrow(test2))

################################## Comparing with base model

# Updating the prediction to say that Customer will not Purchase

test2$Purchase[test2$PBRAND] <- 0


Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm18 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm18
##   FP   TP   TN   FN 
##    0    0 2729  182 
## attr(,"negative")
## [1] "Not Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm18)
##       acc      sens      spec       ppv       npv       lor 
## 0.9374785 0.0000000 1.0000000       NaN 0.9374785       NaN 
## attr(,"negative")
## [1] "Not Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.06252147

# Strategy 19 - No of Car Policy

# Resetting the original training and test data - just to be sure

train <- train.ori
test  <- test.ori
test2 <-test

# Also resetting the test2 data with no one purchased ZeroR strategy

test2$Purchase <- rep(0, nrow(test2))

################################## Comparing with base model

# Updating the prediction to say that Customer will Purchase

test2$Purchase[test2$APERSAUT] <- 1


Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm19 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Purchased")
cm19
##   FP   TP   TN   FN 
##  182 2724    0    5 
## attr(,"negative")
## [1] "Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm19)
##       acc      sens      spec       ppv       npv       lor 
## 0.9357609 0.9981678 0.0000000 0.9373710 0.0000000      -Inf 
## attr(,"negative")
## [1] "Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.06423909

# Strategy 20 - No of Car Policy

# Resetting the original training and test data - just to be sure

train <- train.ori
test  <- test.ori
test2 <-test

# Also resetting the test2 data with no one purchased ZeroR strategy

test2$Purchase <- rep(0, nrow(test2))

################################## Comparing with base model

################ Updating the prediction to say that Customer will not Purchase

test2$Purchase[test2$APERSAUT] <- 0


Our.Prediction=factor(as.factor(test2$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))
Actual.Outcome=factor(as.factor(test$Purchase), c(0, 1), labels = c("Not Purchased", "Purchased"))

# Confusion matrix # cm(actual,predicted)

cm20 = confusionMatrix(Actual.Outcome,Our.Prediction, negative = "Not Purchased")
cm20
##   FP   TP   TN   FN 
##    0    0 2729  182 
## attr(,"negative")
## [1] "Not Purchased"

# corresponding accuracy, sensitivity etc.

diagnosticErrors(cm20)
##       acc      sens      spec       ppv       npv       lor 
## 0.9374785 0.0000000 1.0000000       NaN 0.9374785       NaN 
## attr(,"negative")
## [1] "Not Purchased"

# Computing the classification error

ce(Actual.Outcome,Our.Prediction)
## [1] 0.06252147
##Model Cost Summary
** FINAL OBSERVATION OF THE PROJECT AND ITS MODEL**
1. More than 90% of our observations were responded that they don’t hold Caravan Policy 2. We tried to do the prediction assuming no one will purchase the policy as majority of observations doesnt hold caravan insurance 3. We have used 4 models with more than 12 variables 4. As predominantely the data belongs to one prediction variable there is not much variance in each model accuracy we used for prediction 5. We also tried predicting the another way i.e. customer will purchase and the accuracy has seen notable level of improvement which we have given for ref in the last few models.