WQD7004 Group Assignment
Group Members
- Khairun Nadzirah Binti Abdul Karim (22066330)
- Muhammad Shahzad Rafiq (S2150889)
- Sadman Chowdhury (S2199546)
- Yin QiXiang (S2150692)
- Xin Dong (22060696)
Title
Credit Card Customer Churn Detection Using Machine Learning
Algorithms 
Dataset
Credit Card Fraud Data
https://data.world/vlad/credit-card-fraud-detection
Introduction
The rapid growth of the banking industry has allowed consumers to be more discerning about the banks they want to maintain relationships with. Thus, customer retention has become a significant concern for many financial institutions. One particular area where customer retention is particularly significant is in the realm of credit cards. High churn rates, the rate at which customers stop doing business with an entity, can lead to significant revenue losses and higher acquisition costs for new customers. This project aims to predict credit card customer churn, to help banks identify and retain customers at risk of churning.
Problem Statement
Customer churn in the banking sector, particularly in credit cards, is a persistent issue. Predicting churn can be a complex task due to the multitude of factors that can influence a customer’s decision to leave, including customer service quality, better offerings from competitors, changes in customer financial circumstances, and more. Despite the advent of advanced data analytics techniques, many banks still struggle to predict and mitigate customer churn effectively. This project will focus on this problem, attempting to develop a model that can accurately predict customer churn and thus provide valuable insights to help banks retain their valuable credit card customers.
Research Objective
- To understand the factors that contribute to credit card customer churn.
- To develop a predictive model for credit card customer churn.
- To provide recommendations for customer churn reduction strategies.
Research Question
- What are the key factors influencing credit card customer churn?
- How accurately can we predict credit card customer churn?
- What strategies can banks implement to reduce churn rates among credit card customers?
Dataset
# Loading necessary libraries
library(readxl)
library(ggplot2)
library(dplyr)
library(corrplot)
library(hexbin)
library(plyr)
library(tidyr)
library(purrr)
library(gridExtra)
library(ggrepel)
library(pastecs)
library(caret)
#library(ROSE)
library(randomForest)
library(e1071)
library(rpart)
library(rpart.plot)c_data <- read.csv('BankChurners.csv')
head(c_data, 3)## CLIENTNUM Attrition_Flag Customer_Age Gender Dependent_count
## 1 768805383 Existing Customer 45 M 3
## 2 818770008 Existing Customer 49 F 5
## 3 713982108 Existing Customer 51 M 3
## Education_Level Marital_Status Income_Category Card_Category Months_on_book
## 1 High School Married $60K - $80K Blue 39
## 2 Graduate Single Less than $40K Blue 44
## 3 Graduate Married $80K - $120K Blue 36
## Total_Relationship_Count Months_Inactive_12_mon Contacts_Count_12_mon
## 1 5 1 3
## 2 6 1 2
## 3 4 1 0
## Credit_Limit Total_Revolving_Bal Avg_Open_To_Buy Total_Amt_Chng_Q4_Q1
## 1 12691 777 11914 1.335
## 2 8256 864 7392 1.541
## 3 3418 0 3418 2.594
## Total_Trans_Amt Total_Trans_Ct Total_Ct_Chng_Q4_Q1 Avg_Utilization_Ratio
## 1 1144 42 1.625 0.061
## 2 1291 33 3.714 0.105
## 3 1887 20 2.333 0.000
## Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_1
## 1 9.3448e-05
## 2 5.6861e-05
## 3 2.1081e-05
## Naive_Bayes_Classifier_Attrition_Flag_Card_Category_Contacts_Count_12_mon_Dependent_count_Education_Level_Months_Inactive_12_mon_2
## 1 0.99991
## 2 0.99994
## 3 0.99998Data Cleaning
# Remove duplicates
c_data <- unique(c_data)# Check for null values in each column
null_counts <- sapply(c_data, function(x) sum(is.na(x)))# Drop unnecessary columns
c_data <- c_data[, -c(1, 22, 23)]EDA
Distribution of Customer Age
The distribution of customer ages in the dataset follows a fairly normal distribution. The box plot provides an overview of the median, quartiles, and any potential outliers in the age variable. The histogram illustrates the count of customers in each age group.
Next, we will perform similar EDA analysis for other variables in the dataset:
Distribution of Gender
More samples of females in our dataset are compared to males, but the
percentage of difference is not that significant, so we can say that
genders are uniformly distributed.
Distribution of Education Level
Correlation between Numeric Variables
The correlation matrix provides an overview of the relationships between the numeric variables in the dataset. It helps identify any strong positive or negative correlations between variables.
Scatter Plot: Total_Trans_Amt vs. Total_Trans_Ct
The scatter plot showcases the relationship between the total transaction amount and the total transaction count. It helps visualize any patterns or trends between these two variables.
Summary Statistics
The summary statistics provide a comprehensive overview of the numerical variables in the dataset. It includes measures such as mean, median, standard deviation, minimum, maximum, and various percentiles.
This EDA analysis provides insights into the distribution, relationships, and summary statistics of the key variables in the dataset. Further exploratory analyses can be conducted for other variables as per the project requirements.
Distribution of dependent counts:
The distribution of dependent counts is fairly normally distributed with a slight right skew.
Distribution of months the customer is part of the bank:
Proportion of Education Levels:
If most of the customers with unknown education status lack any education, we can state that more than 70% of the customers have a formal education level. About 35% have a higher level of education.
Kurtosis of Months on book features:
## [1] "Kurtosis of Months on book features is: 0.398638886235621"Distribution of the Total Transaction Amount (Last 12 months):
Data Processing
# Identify the column names of categorical variables and factors
categorical_columns <- sapply(c_data, is.character)
categorical_column_names <- names(c_data[categorical_columns])
# Print the column names of categorical variables and factors# Convert values of Attrition_Flag to 0 and 1
c_data$Attrition_Flag <- ifelse(c_data$Attrition_Flag == "Existing Customer", 0, 1)## 'data.frame': 10127 obs. of 20 variables:
## $ Attrition_Flag : num 0 0 0 0 0 0 0 0 0 0 ...
## $ Customer_Age : int 45 49 51 40 40 44 51 32 37 48 ...
## $ Gender : chr "M" "F" "M" "F" ...
## $ Dependent_count : int 3 5 3 4 3 2 4 0 3 2 ...
## $ Education_Level : chr "High School" "Graduate" "Graduate" "High School" ...
## $ Marital_Status : chr "Married" "Single" "Married" "Unknown" ...
## $ Income_Category : chr "$60K - $80K" "Less than $40K" "$80K - $120K" "Less than $40K" ...
## $ Card_Category : chr "Blue" "Blue" "Blue" "Blue" ...
## $ Months_on_book : int 39 44 36 34 21 36 46 27 36 36 ...
## $ Total_Relationship_Count: int 5 6 4 3 5 3 6 2 5 6 ...
## $ Months_Inactive_12_mon : int 1 1 1 4 1 1 1 2 2 3 ...
## $ Contacts_Count_12_mon : int 3 2 0 1 0 2 3 2 0 3 ...
## $ Credit_Limit : num 12691 8256 3418 3313 4716 ...
## $ Total_Revolving_Bal : int 777 864 0 2517 0 1247 2264 1396 2517 1677 ...
## $ Avg_Open_To_Buy : num 11914 7392 3418 796 4716 ...
## $ Total_Amt_Chng_Q4_Q1 : num 1.33 1.54 2.59 1.4 2.17 ...
## $ Total_Trans_Amt : int 1144 1291 1887 1171 816 1088 1330 1538 1350 1441 ...
## $ Total_Trans_Ct : int 42 33 20 20 28 24 31 36 24 32 ...
## $ Total_Ct_Chng_Q4_Q1 : num 1.62 3.71 2.33 2.33 2.5 ...
## $ Avg_Utilization_Ratio : num 0.061 0.105 0 0.76 0 0.311 0.066 0.048 0.113 0.144 ...# Create a formula for one-hot encoding
formula <- as.formula(paste("factor(Attrition_Flag) ~", paste(categorical_cols, collapse = "+")))
# Create dummy variables using dummyVars
dummy_data <- predict(dummyVars(formula, data = c_data), newdata = c_data)# Combine numerical and one-hot encoded data
combined_data <- cbind(c_data[, !(names(c_data) %in% categorical_cols)], dummy_data)# Split the data into training and testing sets
set.seed(42)
train_indices <- createDataPartition(combined_data$Attrition_Flag, p = 0.7, list = FALSE)
train_data <- combined_data[train_indices, ]
test_data <- combined_data[-train_indices, ]# Separate predictors (x) and target variable (y) in the training and testing sets
X_train <- train_data[, !(names(train_data) %in% "Attrition_Flag")]
y_train <- train_data$Attrition_Flag
X_test <- test_data[, !(names(test_data) %in% "Attrition_Flag")]
y_test <- test_data$Attrition_FlagMachine Learning Modeling
# Random Forest Classifier
rf_model <- randomForest(x = X_train, y = as.factor(y_train), class.factors = levels(as.factor(y_train)))# Train the SVM model
svm_model <- svm(x = X_train, y = as.factor(y_train))# Train the Decision Tree model
dt_model <- rpart(y_train ~ ., data = X_train, method = "class")Performance Metrics
Random Forest Classifier
# Make predictions on the test set
rf_predictions <- predict(rf_model, X_test)
# Convert y_test to have the same levels as rf_predictions
y_test <- factor(y_test, levels = levels(rf_predictions))
# Calculate accuracy and confusion matrix
rf_accuracy <- sum(rf_predictions == y_test) / length(y_test)
rf_confusion <- confusionMatrix(rf_predictions, y_test)
# Print accuracy and confusion matrix
print(paste("Random Forest Accuracy:", rf_accuracy))## [1] "Random Forest Accuracy: 0.953258722843976"print("Random Forest Confusion Matrix:")## [1] "Random Forest Confusion Matrix:"print(rf_confusion$table)## Reference
## Prediction 0 1
## 0 2509 119
## 1 23 387SVM Model
# SVM
# Make predictions on the test set
svm_predictions <- predict(svm_model, X_test)
# Evaluate the model performance
svm_accuracy <- sum(svm_predictions == y_test) / length(y_test)
# Create the confusion matrix
svm_confusion <- confusionMatrix(svm_predictions, y_test)
print(paste("SVM Accuracy:", svm_accuracy))## [1] "SVM Accuracy: 0.910138248847926"print("SVM Confusion Matrix:")## [1] "SVM Confusion Matrix:"print(svm_confusion$table)## Reference
## Prediction 0 1
## 0 2490 231
## 1 42 275Decision Tree model
# Decision Tree
# Predict class labels on the test set
dt_predictions <- predict(dt_model, newdata = X_test, type = "class")
# Evaluate the model performance
dt_accuracy <- sum(dt_predictions == y_test) / length(y_test)
dt_confusion <- confusionMatrix(dt_predictions, y_test)
print(paste("Decision Tree Accuracy:", dt_accuracy))## [1] "Decision Tree Accuracy: 0.929229756418697"print("Decision Tree Confusion Matrix:")## [1] "Decision Tree Confusion Matrix:"print(dt_confusion$table)## Reference
## Prediction 0 1
## 0 2449 132
## 1 83 374Performance Metrics
## Model Accuracy Precision Recall F1_Score
## 1 Random Forest 0.9532587 0.9547184 0.9909163 0.9724806
## 2 SVM 0.9101382 0.9151047 0.9834123 0.9480297
## 3 Decision Tree 0.9292298 0.9488570 0.9672196 0.9579503# Reshape the performance data frame for plotting
performance_long <- reshape2::melt(performance, id.vars = "Model")
# Grouped Bar Chart for Performance Metrics
bar_chart <- ggplot(performance_long, aes(x = Model, y = value, fill = variable)) +
geom_bar(stat = "identity", position = "dodge") +
labs(x = "Model", y = "Performance", title = "Performance Metrics") +
scale_fill_manual(values = c(Accuracy = "blue", Precision = "red", Recall = "green", F1_Score = "purple")) +
theme_minimal() +
theme(legend.position = "right")
print(bar_chart)
# performance of the best model (Roc curve)
library(pROC)## Type 'citation("pROC")' for a citation.##
## Attaching package: 'pROC'## The following objects are masked from 'package:stats':
##
## cov, smooth, var# Calculate ROC curve
rf_predictions_char <- as.numeric(as.character(rf_predictions))
rf_roc <- roc(response = as.numeric(y_test), predictor = rf_predictions_char)## Setting levels: control = 1, case = 2## Setting direction: controls < cases# Plot ROC curve
plot(rf_roc, main = "ROC Curve", print.thres = "best", legacy.axes = TRUE)
Discussion
In performance metrics we using the different kinds of methodology to training dataset and test dataset Including Random Forest Classifier, SVM Model, Decision Tree model. Through the different model performance metrics we could know Random Forest is the highest in accuracy rate. Besides that, in precision, recall and F1 score the Random Forest performance better than other models. Therefore, Random Forest model will adopted as our first choice.
Conclusion
- There are 16.07% of customers who have churned.
- The proportion of gender count is almost equally distributed (52.9%
male and 47.1%) compare to proportion of existing and attributed
customer count (83.9% and 16.1%) which is highly imbalanced
- The proportion of attrited customers by gender there are 14.4% more
male than female who have churned
- Customers who have churned are highly educated - A high proportion
of education level of attrited customer is Graduate level (29.9%),
followed by Post-Graduate level (18.8%)
- A high proportion of marital status of customers who have churned is
Married (43.6%), followed by Single (41.1%) compared to Divorced (7.4%)
and Unknown (7.9%) status - Marital stuats of the attributed customers
are highly clustered in Married status and Single
- As you can see from the proportion of income category of attrited customer, it is highly concentrated around $60K - $80K income (37.6%), followed by Less than $40K income (16.7%) compare to attrited customers with higher annual income of 80K-120K(14.9%) and over $120K + (11.5%). I assume that customers with higher income doesn’t likely to leave their credit card services than meddle-income customer