Customer_Churn.Rmd

---
title: "Telco Customer Churn"
author: "Thanasis Zoumpekas"
output:
  html_document:
    df_print: paged
---

# Data Acquisition

The dataset downloaded from IBM sample datasets. You can find more on link : https://www.ibm.com/communities/analytics/watson-analytics-blog/guide-to-sample-datasets/

### Telco Customer Churn Dataset

This data set provides information to predict behavior to retain customers.

A **telecommunications company** is concerned about the number of customers leaving their landline business for cable competitors. They need to understand who is leaving. 

The data set includes information about:

    -Customers who left within the last month – the column is called Churn
    -Services that each customer has signed up for – phone, multiple lines, internet, online security, online backup, device protection, tech support, and streaming TV and movies
    -Customer account information – how long they’ve been a customer, contract, payment method, paperless billing, monthly charges, and total charges
    -Demographic info about customers – gender, age range, and if they have partners and dependents
    


```{r}

library(readr)
library(tidyr)
library(tidyverse) 
library(MASS)
library(car)
library(e1071)
library(caret)
library(cowplot)
library(caTools)
library(pROC)


df <- read.csv('/home/thzou/Projects/B.I./Customer_Churn/WA_Fn-UseC_-Telco-Customer-Churn.csv')
head(df)

```


```{r}
summary(df)
```

# Data Pre-processing


## Data Exploration

We will check for na values, i.e. missing data from our dataframe. 

```{r}

options(repr.plot.width = 6, repr.plot.height = 4)
missing_data <- df %>% summarise_all(funs(sum(is.na(.))/n()))
missing_data <- gather(missing_data, key = "variables", value = "percent_missing")
ggplot(missing_data, aes(x = reorder(variables, percent_missing), y = percent_missing)) +
geom_bar(stat = "identity", fill = "red", aes(color = I('white')), size = 0.3)+
xlab('variables')+
coord_flip()+ 
theme_bw()


```

```{r}

cat ("Number of missing values :", sum(is.na(df))) #number of missing values
dfnew = na.omit(df)
cat ("\nNumber of missing values (after removing the NAs):", sum(is.na(dfnew))) #double check for missing values

head(dfnew)
```

SeniorCitizen is in 'int' form, that can be changed to categorical.

```{r}


dfnew <- dfnew[complete.cases(dfnew),]

dfnew$SeniorCitizen <- as.factor(ifelse(dfnew$SeniorCitizen==1, 'YES', 'NO'))

```




We will check how many unique customers we have. As we can see we have 7032 unique customers.
```{r}
length(unique(dfnew$customerID)) 
```

We will drop the column customerID, because it is unique for each observation. Thus, this variable does not provide us with additional valuable information.

```{r}
dfnew = subset(dfnew, select = -c(customerID)) #drop the customerID
head(dfnew)
```


```{r}


options(repr.plot.width = 6, repr.plot.height = 4)
dfnew %>% 
group_by(Churn) %>% 
summarise(Count = n())%>% 
mutate(percent = prop.table(Count)*100)%>%
ggplot(aes(reorder(Churn, -percent), percent), fill = Churn)+
geom_col(fill = c("#FC4E07", "#E7B800"))+
geom_text(aes(label = sprintf("%.2f%%", percent)), hjust = 0.01,vjust = -0.5, size =3)+ 
theme_bw()+  
xlab("Churn") + 
ylab("Percent")+
ggtitle("Churn Percent")


```


The graph above shows that 26% of customers or subscribers stop doing business with the company.

Below we can see the percentage of customer churn, according to each feature.
```{r}

theme1 <- theme_bw()+ theme(axis.text.x = element_text(angle = 0, hjust = 1, vjust = 0.5),legend.position="none")
theme2 <- theme_bw()+ theme(axis.text.x = element_text(angle = 90, hjust = 1, vjust = 0.5),legend.position="none")


#options(repr.plot.width = 14, repr.plot.height = 10)
plot_grid(ggplot(dfnew, aes(x=gender,fill=Churn))+ geom_bar()+ theme2, 
          ggplot(dfnew, aes(x=SeniorCitizen,fill=Churn))+ geom_bar(position = 'fill')+theme2,
          ggplot(dfnew, aes(x=Partner,fill=Churn))+ geom_bar(position = 'fill')+theme2,
          ggplot(dfnew, aes(x=Dependents,fill=Churn))+ geom_bar(position = 'fill')+theme2,
          ggplot(dfnew, aes(x=PhoneService,fill=Churn))+ geom_bar(position = 'fill')+theme2,
          ggplot(dfnew, aes(x=MultipleLines,fill=Churn))+ geom_bar(position = 'fill')+theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 14)),
          align = "h")
```

```{r}

#options(repr.plot.width = 12, repr.plot.height = 8)
plot_grid(ggplot(dfnew, aes(x=InternetService,fill=Churn))+ geom_bar(position = 'fill')+ theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)), 
          ggplot(dfnew, aes(x=OnlineSecurity,fill=Churn))+ geom_bar(position = 'fill')+theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)),
          ggplot(dfnew, aes(x=OnlineBackup,fill=Churn))+ geom_bar(position = 'fill')+theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)),
          ggplot(dfnew, aes(x=DeviceProtection,fill=Churn))+ geom_bar(position = 'fill')+theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)),
          ggplot(dfnew, aes(x=TechSupport,fill=Churn))+ geom_bar(position = 'fill')+theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)),
          ggplot(dfnew, aes(x=StreamingTV,fill=Churn))+ geom_bar(position = 'fill')+theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)),
          align = "h")


```


```{r}

plot_grid(ggplot(dfnew, aes(x=StreamingMovies,fill=Churn))+ 
          geom_bar(position = 'fill')+ theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)), 
          ggplot(dfnew, aes(x=Contract,fill=Churn))+ 
          geom_bar(position = 'fill')+theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)),
          ggplot(dfnew, aes(x=PaperlessBilling,fill=Churn))+ 
          geom_bar(position = 'fill')+theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)),
          ggplot(dfnew, aes(x=PaymentMethod,fill=Churn))+
          geom_bar(position = 'fill')+theme2+
          scale_x_discrete(labels = function(x) str_wrap(x, width = 10)),
          align = "h")


```


```{r}

ggplot(dfnew, aes(x=Churn, y=tenure, fill=Churn)) +
    geom_boxplot(alpha=0.4) +
    stat_summary(fun.y=mean, geom="point", shape=20, size=10, color="red", fill="red") +
    theme(legend.position="right") +
    scale_fill_brewer(palette="Set2") +
    xlab(" ")

 


ggplot(dfnew, aes(x=Churn, y=MonthlyCharges, fill=Churn)) +
    geom_boxplot(alpha=0.4) +
    stat_summary(fun.y=mean, geom="point", shape=20, size=10, color="red", fill="red") +
    theme(legend.position="right") +
    scale_fill_brewer(palette="Set1") +
    xlab(" ")

 


ggplot(dfnew, aes(x=Churn, y=TotalCharges, fill=Churn)) +
    geom_boxplot(alpha=0.4) +
    stat_summary(fun.y=mean, geom="point", shape=20, size=10, color="red", fill="red") +
    theme(legend.position="right") +
    scale_fill_brewer(palette="Set3") +
    xlab(" ")

 
```



The red dot on the graphs above indicates the mean (average value).

We will check the correlation between continuous variables of the dataset.
```{r}
library(mlbench)
library(caret)
library(corrplot)

# calculate correlation matrix
correlationMatrix <- cor(dfnew[,c('tenure','MonthlyCharges','TotalCharges')])
# summarize the correlation matrix
print(correlationMatrix)
# find attributes that are highly corrected (ideally >0.75)
highlyCorrelated <- findCorrelation(correlationMatrix, cutoff=0.5)
# print indexes of highly correlated attributes
print(highlyCorrelated)


corrplot(correlationMatrix, method="number")


```



We have only 3 continuous variables so we dont need a heatmap to represent the correlation matrix. As we can see Total Charges are highly correlated (0.83) with tenure. We can also observe that Monthly Charges are also correlated with Total Charges (0.65). That is somehow natural, because they are both charges.

## Data Preparation

**Redefine the categorical features-variables**

According to Data Exploration section, we observe that there are categorical features, labeled 'No', 'No Internet Service' or 'No Phone Service'. In order to make these variables more useful, we redefine these values as 'No' and clean these features.


```{r}
dfnew <- data.frame(lapply(dfnew, function(x) {
                  gsub("No internet service", "No", x)}))

dfnew <- data.frame(lapply(dfnew, function(x) {
                  gsub("No phone service", "No", x)}))
```


**Standardization of continuous variables**

```{r}
num_columns <- c("tenure", "MonthlyCharges", "TotalCharges")
dfnew[num_columns] <- sapply(dfnew[num_columns], as.numeric)

telco_int <- dfnew[,c("tenure", "MonthlyCharges", "TotalCharges")]
telco_int <- data.frame(scale(telco_int))
```


**Creating derived features**

In order to create derived feature from tenure(which is in months), we create different bins of tenure, labeled as follows, '0-1 year', '2-3 years', '3-4 years' etc.

```{r}
#max(telco$tenure)
#min(telco$tenure)
dfnew <- mutate(dfnew, tenure_bin = tenure)

dfnew$tenure_bin[dfnew$tenure_bin >=0 & dfnew$tenure_bin <= 12] <- '0-1 year'
dfnew$tenure_bin[dfnew$tenure_bin > 12 & dfnew$tenure_bin <= 24] <- '1-2 years'
dfnew$tenure_bin[dfnew$tenure_bin > 24 & dfnew$tenure_bin <= 36] <- '2-3 years'
dfnew$tenure_bin[dfnew$tenure_bin > 36 & dfnew$tenure_bin <= 48] <- '3-4 years'
dfnew$tenure_bin[dfnew$tenure_bin > 48 & dfnew$tenure_bin <= 60] <- '4-5 years'
dfnew$tenure_bin[dfnew$tenure_bin > 60 & dfnew$tenure_bin <= 72] <- '5-6 years'

dfnew$tenure_bin <- as.factor(dfnew$tenure_bin)
```


After observing the distribution of data in each tenure bin, we found that maximum number of customers have a tenure of either 0-1 years and followed by 5-6 years.


```{r}
ggplot(dfnew, aes(tenure_bin, fill = tenure_bin)) + geom_bar()+ theme1
```



**Creating Dummy Variables**


```{r}

telco_cat <- dfnew[,-c(5,18,19)]

#Creating Dummy Variables
dummy<- data.frame(sapply(telco_cat,function(x) data.frame(model.matrix(~x-1,data =telco_cat))[,-1]))

head(dummy)


```




**Creating the final dataset by combining the numeric and dummy data frames.**

```{r}
#Combining the data
df_final <- cbind(telco_int,dummy)
head(df_final)

```

```{r}

col_names <- names(df_final[,-c(1:3)])
df_final[,col_names] <- lapply(df_final[,col_names] , factor)

```


## Rank Features By Importance

The importance of features can be estimated from data by building a model. Some methods like decision trees have a built in mechanism to report on variable importance. For other algorithms, the importance can be estimated using a ROC curve analysis conducted for each attribute.

Below we will use the Learning Vector Quantization (LVQ) model. The varImp is then used to estimate the variable importance, which is printed and plotted. It shows that the glucose, mass and age attributes are the top 3 most important attributes in the dataset and the insulin attribute is the least important.

```{r}
set.seed(6)
library(e1071)
library(mlbench)
library(caret)
library(doParallel)

cl <- makeCluster(4)
registerDoParallel(cl)

# prepare training scheme
control <- trainControl(method="repeatedcv", number=10, repeats=3)
# train the model
model_lvq <- train(Churn~., data=df_final, method="lvq", trControl=control)
# estimate variable importance
importance_lvq <- varImp(model_lvq, scale=FALSE)
stopCluster(cl)
# summarize importance
print(importance_lvq)
# plot importance
plot(importance_lvq)
```




```{r}
set.seed(7)
library(e1071)
library(mlbench)
library(caret)
library(doParallel)

cl <- makeCluster(4)
registerDoParallel(cl)

# prepare training scheme
control <- trainControl(method="repeatedcv", number=10, repeats=3)
# train the model
model_ada <- train(Churn~., data=df_final, method="ada", trControl=control)
# estimate variable importance
importance_ada <- varImp(model_ada, scale=FALSE)
stopCluster(cl)
# summarize importance
print(importance_ada)
# plot importance
plot(importance_ada)
```

```{r}
set.seed(8)
library(e1071)
library(mlbench)
library(caret)
library(doParallel)

cl <- makeCluster(4)
registerDoParallel(cl)

# prepare training scheme
control <- trainControl(method="repeatedcv", number=10, repeats=3)
# train the model
model_svm <- train(Churn~., data=df_final, method="svmRadial", trControl=control)
# estimate variable importance
importance_svm <- varImp(model_svm, scale=FALSE)
stopCluster(cl)
# summarize importance
print(importance_svm)
# plot importance
plot(importance_svm)
```



## Feature Selection

Automatic feature selection methods can be used to build many models with different subsets of a dataset and identify those attributes that are and are not required to build an accurate model.

A popular automatic method for feature selection provided by the caret R package is called Recursive Feature Elimination or RFE.

The example below provides an example of the RFE method on the Pima Indians Diabetes dataset. A Random Forest algorithm is used on each iteration to evaluate the model. The algorithm is configured to explore all possible subsets of the attributes. All 8 attributes are selected in this example, although in the plot showing the accuracy of the different attribute subset sizes, we can see that just 4 attributes gives almost comparable results.

```{r}
# ensure the results are repeatable
set.seed(7)
# load the library
library(mlbench)
library(caret)
library(doParallel)

y = df_final$Churn
x = subset(df_final, select = -c(Churn)) 

# define the control using a random forest selection function
cl <- makeCluster(4)
registerDoParallel(cl)
control <- rfeControl(functions=rfFuncs, method="cv", number=10, repeats=3)
# run the RFE algorithm
results <- rfe(x, y, rfeControl=control)
stopCluster(cl)
# summarize the results
print(results)
# list the chosen features
predictors(results)
# plot the results
plot(results, type=c("g", "o"))
```


## Model Building

We will try to predict if a customer is going to leave of not (i.e. customer churn prediction).

### Logistic Regression


First of all we will split the data in 65% training and 35% validation set.
```{r}

library(e1071)
library(mlbench)
library(caret)
library(doParallel)


#Splitting the data
set.seed(123)
indices = sample.split(df_final$Churn, SplitRatio = 0.65)
train = df_final[indices,]
validation = df_final[!(indices),]

cl <- makeCluster(4)
registerDoParallel(cl)

#Build the first model using all variables
model_1 = glm(Churn ~ ., data = train, family = "binomial")
summary(model_1)

stopCluster(cl)


```




**stepAIC** is a iterative process of adding or removing variables, in order to get a subset of variables that gives the best performing model. We will use this for variable selection.



```{r}
model_2 <- stepAIC(model_1, direction="both",trace = 0)
```


```{r}
summary(model_2)
```



We use variance inflation factor (vif) to get rid of redundant predictors or the variables that have high multicollinearity between them. Multicollinearity exists when two or more predictor variables are highly related to each other and then it becomes difficult to understand the impact of an independent variable on the dependent variable.

The Variance Inflation Factor(VIF) is used to measure the multicollinearity between predictor variables in a model. A predictor having a VIF of 2 or less is generally considered safe and it can be assumed that it is not correlated with other predictor variables. Higher the VIF, greater is the correlation of the predictor variable w.r.t other predictor variables. However, Predictors with high VIF may have high p-value(or highly significant), hence, we need to see the significance of the Predictor variable before removing it from our model.


```{r}
vif(model_2)
```


**Removing variables-features due to high p-value ( > 0.05 alpha level)**


```{r}

model_3 <-glm(formula = Churn ~ tenure + MonthlyCharges + SeniorCitizen + 
    InternetService.xFiber.optic + InternetService.xNo + 
    OnlineSecurity + OnlineBackup + TechSupport + 
    StreamingTV + Contract.xOne.year + Contract.xTwo.year + PaperlessBilling + 
    PaymentMethod.xElectronic.check + tenure_bin.x1.2.years + 
    tenure_bin.x5.6.years, family = "binomial", data = train)

summary(model_3)

vif(model_3)
```

**Removing StreamingTV feature**

```{r}
model_4 <-glm(formula = Churn ~ tenure + MonthlyCharges + SeniorCitizen + 
    InternetService.xFiber.optic + InternetService.xNo + 
    OnlineSecurity + OnlineBackup + TechSupport + 
    Contract.xOne.year + Contract.xTwo.year + PaperlessBilling + 
    PaymentMethod.xElectronic.check + tenure_bin.x1.2.years + 
    tenure_bin.x5.6.years, family = "binomial", data = train)

summary(model_4)

vif(model_4)
```


## Model Evaluation

### Logistic Regression

```{r}

final_model <- model_4 #keep the final model


pred <- predict(final_model, type = "response", newdata = validation[,-24])
summary(pred)
validation$prob <- pred

# Using probability cutoff of 50%.

pred_churn <- factor(ifelse(pred >= 0.50, "Yes", "No"))
actual_churn <- factor(ifelse(validation$Churn==1,"Yes","No"))
table(actual_churn,pred_churn)

```


```{r}
cutoff_churn <- factor(ifelse(pred >=0.50, "Yes", "No"))
conf_final <- confusionMatrix(cutoff_churn, actual_churn, positive = "Yes")
accuracy <- conf_final$overall[1]
sensitivity <- conf_final$byClass[1]
specificity <- conf_final$byClass[2]

round(accuracy,digits=2)
round(sensitivity,digits=2)
round(specificity,digits=2)

```


We can observe that, by using a cutoff of 0.50, we are getting good accuracy and specificity, but the sensitivity is nearly above average. Thus, we need to find the optimal probalility cutoff which will give maximum accuracy, sensitivity and specificity.



```{r}

perform_fn <- function(cutoff) 
{
  predicted_churn <- factor(ifelse(pred >= cutoff, "Yes", "No"))
  conf <- confusionMatrix(predicted_churn, actual_churn, positive = "Yes")
  accuray <- conf$overall[1]
  sensitivity <- conf$byClass[1]
  specificity <- conf$byClass[2]
  out <- t(as.matrix(c(sensitivity, specificity, accuray))) 
  colnames(out) <- c("sensitivity", "specificity", "accuracy")
  return(out)
}

summary(pred)
s = seq(0.01,0.80,length=100)
OUT = matrix(0,100,3)

for(i in 1:100)
{
  OUT[i,] = perform_fn(s[i])
} 

plot(s, OUT[,1],xlab="Cutoff",ylab="Value",cex.lab=1.5,cex.axis=1.5,ylim=c(0,1),
     type="l",lwd=2,axes=FALSE,col=2)
axis(1,seq(0,1,length=5),seq(0,1,length=5),cex.lab=1.5)
axis(2,seq(0,1,length=5),seq(0,1,length=5),cex.lab=1.5)
lines(s,OUT[,2],col="darkgreen",lwd=2)
lines(s,OUT[,3],col=4,lwd=2)
box()
legend("bottom",col=c(2,"darkgreen",4,"darkred"),text.font =3,inset = 0.02,
       box.lty=0,cex = 0.8, 
       lwd=c(2,2,2,2),c("Sensitivity","Specificity","Accuracy"))
abline(v = 0.32, col="red", lwd=1, lty=2)
axis(1, at = seq(0.1, 1, by = 0.1))


optimal_cutoff <- s[which(abs(OUT[,1]-OUT[,2])<0.01)]

cat("\n\nOptimal Cutoff : ", optimal_cutoff)
```

We will calculate again the accuracy, sensitivity and specificity, by using optimal probalility cutoff.

```{r}
cutoff_churn <- factor(ifelse(pred >= optimal_cutoff, "Yes", "No"))
conf_final <- confusionMatrix(cutoff_churn, actual_churn, positive = "Yes")
accuracy <- conf_final$overall[1]
sensitivity <- conf_final$byClass[1]
specificity <- conf_final$byClass[2]

accuracy = round(accuracy,digits=2)
sensitivity = round(sensitivity,digits=2)
specificity = round(specificity,digits=2)

```


## Model Summary 

Best logistic regression model : 

```{r}

summary(final_model)

perf_metrics <- data_frame(accuracy,sensitivity,specificity)
perf_metrics
```


We can predict with 75% probability if a customer is going to leave/unsubscribe or stay.