Wednesday 7 May 2014

SUPPORT VECTOR MACHINE (SVM) - Detailed Example on Classification in R



R Code - Bank Subscription Marketing - Classification {SUPPORT VECTOR MACHINE}

 R Code  for Support Vector Machine (SVM)
Data Set:- Bank Marketing https://www.blogger.com/blogger.g?blogID=4359514443959080595#editor/target=post;postID=5426750358330739357
Source of Data Set:-  UCI Repository (http://archive.ics.uci.edu/ml/datasets/Bank+Marketing)
The Code includes the following:-
1. Data Exploration - Missing Values, Outliers
2. Data Visualisation
3.  Correlation Matrix
4. Data Partitioning
5. Package used "e1071"
6. Model Tuning and Plots
7. Confusion Matrix


## The data has been imported using Import Dataset option in R Environment
## The data set can be obtained from http://archive.ics.uci.edu/ml/datasets/Bank+Marketing
## DATASET UNDERSTANDING
head(bank_full) ## Displays first 6 rows for each variable
str(bank_full) ## Describes each variables
summary(bank_full) ## Provides basic statistical information of each variable

## DATA EXPLORATION - Check for Missing Data
## Option 1
is.na(bank_full) ## Displays True for a missing value
## Since it is a large dataset, graphical display of missing values will prove to be easier
##Option 2
require(Amelia)
missmap(bank_full,main="Missing Data - Bank Subscription", col=c("red","grey"),legend=FALSE)

## No red colour stripes are visible. hence no missing values.
##Option 3
summary(bank_full) ## displays missing values if any under every variable

## DATA VISUALISATION
## Use Box plots (Only for continuous variables)- To Check Ouliers
boxplot(bank_full$age~bank_full$subscribed, main=" AGE",ylab="age of customers",xlab="Subscribed")
boxplot(bank_full$balance~bank_full$subscribed, main=" BALANCE",ylab="Balance of customers",xlab="Subscribed")
boxplot(bank_full$lastday~bank_full$subscribed, main=" LAST DAY",ylab="Last day of contact",xlab="Subscribed")
boxplot(bank_full$lastduration~bank_full$subscribed, main="LAST DURATION",ylab="Last duration of contact",xlab="Subscribed")
boxplot(bank_full$numcontacts~bank_full$subscribed, main="NUM CONTACTS",ylab="number of contacts",xlab="Subscribed")
boxplot(bank_full$pdays~bank_full$subscribed, main=" Previous DAYS",ylab="Previous days of contact",xlab="Subscribed")
boxplot(bank_full$pcontacts~bank_full$subscribed, main=" Previous Contacts",ylab="Previous Contacts with customers",xlab="Subscribed")

## Though some outliers are observed in Previous contacts, NumContacts and LastDuration, they have not bee removed keeping their significance into consideration
## Use Histograms (For both continuous and categorical variables)
## These histograms provide details abpout Skewness, Normal Distribution etc
## Function to create histograms for continuous variables with normal curve
bank_Conthist<-function(VarName,NumBreaks,xlab,main,lengthxfit) ## xlab and main should be mentioned under quotes as they are characters
{
  hist(VarName,breaks=NumBreaks,col="yellow",xlab=xlab,main=main)
  xfit<-seq(min(VarName),max(VarName),length=lengthxfit)
  yfit<-dnorm(xfit,mean=mean(VarName),sd=sd(VarName))
  yfit<-yfit*diff(h$mids[1:2])*length(VarName)
  lines(xfit,yfit,col="red",lwd=3)
}
bank_Conthist(bank_full$age,10,"age of customers","AGE",30)
bank_Conthist(bank_full$balance,50,"Balance of customers","Balance",100)

## Balance is more skewed towards to Negative or Zero
bank_Conthist(bank_full$lastday,5,"Last Day of contact","LAst Day",10)
bank_Conthist(bank_full$lastduration,100,"LastDuration of COntact","Last Duration",10)

## Last Duration is more skewed towards 0 to 100 secs.
bank_Conthist(bank_full$numcontacts,30,"Number of Contacts","NUmContacts",20)
## NUmContacts are more skewed towards 1
bank_Conthist(bank_full$pdays,30,"Previous Days of contacts","PDays",20)
## Many were not contacted previously
bank_Conthist(bank_full$pcontacts,20,"Previous Contacts","PContacts",10)
## Since many were not contacted previously, therefore Pcontacts is 0

## Barplots for Categorical Variables
barplot(table(bank_full$job),col="red",main="JOB")
barplot(table(bank_full$marital),col="green",main="Marital")
barplot(table(bank_full$education),col="red",main="Education")
barplot(table(bank_full$creditdefault),col="red",main="Credit Default")

## Since Credit Default is highly skewed towards NO, this shall be removed from further analysis
bank_full[5]<-NULL
str(bank_full)
barplot(table(bank_full$housingloan),col="red",main="Housing Loan")
barplot(table(bank_full$personalloan),col="blue",main="Personal Loan")
barplot(table(bank_full$lastcommtype),col="red",main="Last communication type")
barplot(table(bank_full$lastmonth),col="violet",main="Last Month")
barplot(table(bank_full$poutcome),col="magenta",main="Previous Outcome")

## Correlation Matrix among input (or independent) continuous variables
bank_full.cont<-data.frame(bank_full$age,bank_full$balance,bank_full$lastday,bank_full$lastduration,bank_full$numcontacts,bank_full$pdays,bank_full$pcontacts)
str(bank_full.cont)
cor(bank_full.cont)

## It can be observed that No two variables are highly correlated


 ## Partitioning Data into Train and Test datasets in 70:30
library(caret)
set.seed(1234567)
train2<-createDataPartition(bank_full$subscribed,p=0.7,list=FALSE)
train<-bank_full[train2,]
test<-bank_full[-train2,]
write.table(bank_full,"bank1.csv",sep=",")
str(train)
str(test)



## CLASSIFICATION USING SUPPORT VECTOR MACHINE(SVM) Model
 


install.packages("e1071")
library(e1071)

## TUNING to perform cross-validation to know about optimum cost parameter for SVM modelling. Tune() performs only ten-fold cross validation here.
bank_full.tune<-tune(svm,subscribed~.,data=bank_full,kernel="linear",ranges=list(cost=c(0.01,0.1,1,5,10,100)))
## Diffferent kernels such as Polynomial, Radial Basis and Sigmoid can also be used for tuning.
summary(bank_full.tune)
bestmodal.tune<-bank_full.tune$best.model
summary(bestmodal.tune)

##We get cost=0.01 as optimum one with kernel = linear.
##In the following example, Polynomial function is being used
train.svm<-svm(subscribed~.,train,kernel="polynomial",cost=0.01,scale=TRUE,degree=3,gamma=1)
summary(train.svm)

plot(train.svm,train)  ## Not possible for large data-sets
## Testing data using the trained model
test.svm<-predict(train.svm,test)
## Misclassification Table
table(predict=test.svm,truth=test$subscribed)

1 comment:

  1. 1) There is no variable called subscribed in bank-fill data
    2) In Function to create histograms for continuous variables with normal curve there you have not defined h ?

    ReplyDelete