Using Machine Learning to Predict Customer Churn

Wrangling

Note there are a small number of missing values in the ‘TotalCharges’ observations. My decision here is to predict a mean value for this missing observation and use all available data, rather than ignore this information completely.

It appears the SeniorCitizen observations have unique values of either 0 or 1 and are actually a factor and these formats will be coerced.

# Replace the na total charges with the mean
data <- data  %>% 
  mutate(TotalCharges = if_else(is.na(TotalCharges) == TRUE,
  mean(TotalCharges, na.rm = TRUE), TotalCharges))
missmap(data)
# View the unique SeniorCitizen values and convert to factor
unique(data$SeniorCitizen)
# Coerce 
data <- data %>% 
  mutate(SeniorCitizen = as.factor(if_else(SeniorCitizen == 1, "Yes", "No")))
data <- data %>% 
  mutate(tenure = as.integer(tenure))

Exploration

Now that the dataset is complete let’s gather some some insights through exploratory data analysis (EDA).

The aim here is to understand key insights for why customers may be churning to identify trends, which can provide a basis for future strategy.

# Churn vs internet service
data %>% ggplot(aes(InternetService, TotalCharges, fill = Churn))+
  geom_col()+
  ggtitle("Total Charges by Internet Service")+
  labs(x="Internet Service", y="Total Charges",
  subtitle = "Higher churn in fibre optic internet services")+
  scale_y_continuous(labels = dollar)
# How do the costs stack up?
data %>% ggplot(aes(tenure, TotalCharges, colour=InternetService))+
  geom_point()+
  ggtitle("Total Charges by Plan & Tenure")+
  labs(x="Tenure", y="Total Charges",
  subtitle = "Note the absence of longer term DSL customers")+
  scale_y_continuous(labels = dollar)
# Lets look more closely at this
data %>% filter(InternetService=="DSL") %>%
  ggplot(aes(tenure, MonthlyCharges, colour=Churn))+
  geom_point()+
  geom_smooth(se=FALSE)+
  ggtitle("Tenure vs Monthly Charges vs Churn")+
  labs(x="Tenure", y="Monthly Charges",
  subtitle = "DSL churn occurs early in tenure without charges correlation")
# Is there a connection with dependents? 
data %>% filter(InternetService=="DSL" & Churn == "Yes") %>%
  ggplot(aes(tenure, MonthlyCharges, colour=Dependents))+
  geom_point()+
  geom_smooth(se=FALSE)+
  ggtitle("Churned DSL customers vs Monthly Charges vs Dependents")+
  labs(x="Tenure", y="Monthly Charges",
  subtitle="It appears if you do have dependents, monthly charges are higher")
# Perhaps this churn is related to contract?
data %>% filter(InternetService=="DSL" & Churn == "Yes") %>%
  ggplot(aes(tenure, MonthlyCharges, colour=Contract))+
  geom_point()+
  geom_smooth(se=FALSE)+
  ggtitle("Churned DSL customers vs Tenure vs Plan")+
  labs(x="Tenure", y="Monthly Charges",
  subtitle="The business needs to convert monthly plans to longer terms")
# Customer value
data %>% group_by(Contract, TotalCharges) %>%
  ggplot(aes(Contract, TotalCharges, fill = Contract))+
  geom_col()+
  ggtitle("Total Revenue by Plan")+
  labs(x="Plan", y="Total Revenue",
       subtitle = "Opportunity to Extend Customer Value From Monthly Plan")+
  scale_y_continuous(labels = dollar)
# Facet wrap plot
data %>% ggplot(aes(SeniorCitizen,TotalCharges, fill=Churn))+
  facet_wrap(~PaymentMethod)+
  geom_col()+
  ggtitle("Payment Method vs Senior Citizen")+
  labs(x="Senior Citizen", y="Total Charges",
       subtitle = "Surprised to see customers using mailed payments not senior citizens")+
  scale_y_continuous(labels = dollar)
# Tenure length wrangling
data <- mutate(data, tenure_bin = tenure)
data$tenure_bin[data$tenure_bin >=0 & data$tenure_bin <= 12] <- '0-1 year'
data$tenure_bin[data$tenure_bin > 12 & data$tenure_bin <= 24] <- '1-2 years'
data$tenure_bin[data$tenure_bin > 24 & data$tenure_bin <= 36] <- '2-3 years'
data$tenure_bin[data$tenure_bin > 36 & data$tenure_bin <= 48] <- '3-4 years'
data$tenure_bin[data$tenure_bin > 48 & data$tenure_bin <= 60] <- '4-5 years'
data$tenure_bin[data$tenure_bin > 60 & data$tenure_bin <= 72] <- '5-6 years'
data$tenure_bin <- as.factor(data$tenure_bin)
# Plot tenure length
data %>% ggplot(aes(tenure_bin, fill = tenure_bin)) + 
  geom_bar()+
  ggtitle("Tenure Length by Year")+
  labs(x="Tenure Length in Years",
       y="Number of Customers",
       subtitle = "Again we see the short term monthly plans as an opportunity")+
  scale_fill_discrete(guide=FALSE)
# Plotly object
plota <- data %>% ggplot(aes(tenure, MonthlyCharges, colour = InternetService))+
  geom_point()+
  geom_smooth()+
  ggtitle("Monthly Charges vs Tenure vs Service")+
  labs(x="Tenure", y="Monthly Charges")
ggplotly(plota)

Feature Importance

Now that we have some initial eda lets use machine learning to understand what features are the most predictive for churn and which are not. Boruta uses a powerful randomforest algorithm to calculate importance.

set.seed(1)
# For ease lets remove the customer identification
boruta_data <- data[complete.cases(data[]),] %>%
  select(-customerID)
# Create and print Boruta output 
boruta_output <- Boruta(Churn~., data = boruta_data)
print(boruta_output)
# Tidy and plot the output
plot(boruta_output, xlab = "", xaxt = "n")
lz<-lapply(1:ncol(boruta_output$ImpHistory),function(i)
boruta_output$ImpHistory[is.finite(boruta_output$ImpHistory[,i]),i])
names(lz) <- colnames(boruta_output$ImpHistory)
Labels <- sort(sapply(lz,median))
axis(side = 1,las=2,labels = names(Labels),
at = 1:ncol(boruta_output$ImpHistory), cex.axis = 0.7)
# Get the important attributes withough tentative
getSelectedAttributes(boruta_output, withTentative = F)

Modelling

Now we will create training and test sets, with a caret powered randomforest and an automated machine learning package, h2o to compare results. From a business case perspective, an accurate predictive model could be used to fire targeted offers at customers with predicted churn.

# Remove unimportant features
data <- data %>% select(-gender, -customerID)
# Split the data into training and validation sets
test_index <- createDataPartition(data$Churn, p = .10, list = FALSE)
training <- data[-test_index,]
validation <- data[test_index,]
# Train a random forest model
rf_fit <- train(Churn ~.,
                data = training,
                method = "ranger")
# Test the model
rf_pred <- predict(rf_fit, newdata = validation, na.action = na.pass)
# Table and view the result
rf_result <- confusionMatrix(table(rf_pred,validation$Churn))
rf_result
# Store the accuracy for comparison later
rf.accuracy <- rf_result$overall['Accuracy']

Automated Machine Learning (AML)

Now we’ll use an automated alogorithm that searches for the best fit model including stacked ensembles and compare results.

set.seed(1234)
# Use the h2o package to create a best fit stacked ensemble
h2o.init()
# Convert data to h2o arrays
h2o_training <- as.h2o(training)
h2o_test <- as.h2o(validation)
# Use the power of automated machine learning
aml <- h2o.automl(y="Churn", training_frame = h2o_training,
                  max_runtime_secs = 300)
aml_pred <- h2o.predict(aml@leader, h2o_test)
# Store accuracy and create confusion matrix
perf <- h2o.performance(aml@leader,h2o_test)
perf_cf <- h2o.confusionMatrix(perf)
h2o_acc <- max(h2o.accuracy(perf))
perf_cf
# Compare best performing automated algorithm vs randomforest
overall_results <- data.frame(Method="randomforest", 
                          Accuracy = rf.accuracy)
h2o_results <- data.frame(Method="aml",
                          Accuracy = h2o_acc)
overall_results <- overall_results %>% 
  rbind(h2o_results) %>%  knitr::kable(row.names = FALSE)
overall_results

Conclusion

It is possible to predict churn for this business with a reasonable level of accuracy and the recommendation would be to build a business case to understand the benefit of targeted promotions for churn predicted customers vs the cost of activation. The promotional cost of such an activity could be reduced by targeting customers with a higher probability of churning.

It is clear that AML processes are very powerful also, though I could not have realised this, without having learned the science and programming background required for other methods - hard to believe such powerful tools are open source and freely available. So, if you are looking to model churn for a saas or retail business, R is very capable for this requirement.

If you think your dataset may be too large, this is most unlikely, especially if you use an on disk manipulation method like disk.frame, where the largest file I heard the developer open was about 160 billion rows and it took a couple of minutes to open.