Customer Churn Analysis

This project builds an end-to-end customer churn analysis and prediction pipeline using Python.
It combines multiple data sources (demographics, transactions, customer service, online activity, and churn labels) to:

• Explore key drivers of churn.
• Engineer useful features.
• Build and evaluate a Random Forest classification model.
• Generate plots that are saved into an images/ folder for use in reports and on GitHub.

The analysis is implemented in a single Jupyter notebook:
Customer_Churn_Analysis.ipynb.

---

1. Project Overview

Customer churn is a critical problem for subscription and service-based businesses.
This project answers three main questions:

1. Who is likely to churn?
2. What customer characteristics and behaviours drive churn?
3. How accurately can we predict churn using a supervised ML model?

The workflow covers:

• Data loading from multiple Excel files.
• Exploratory Data Analysis (EDA).
• Feature engineering and encoding.
• Data cleaning and preprocessing.
• Model training (Random Forest classifier).
• Model evaluation and interpretation.
• Exporting visualisations to an images/ folder for reuse.

---

2. Data Sources

The analysis uses five related Excel datasets (not included in the repo by default):

1. Customer Demographics – Customer_Churn_Data_Large_CD.xlsx

   • Example fields: CustomerID, Age, Gender, MaritalStatus, IncomeLevel, etc.

2. Transaction History – Customer_Churn_Data_Large_TH.xlsx

   • Example fields: CustomerID, TransactionID, TransactionDate, AmountSpent, ProductCategory, etc.

3. Customer Service Interactions – Customer_Churn_Data_Large_CS.xlsx

   • Example fields: CustomerID, InteractionID, InteractionType, ResolutionStatus, etc.

4. Online Activity – Customer_Churn_Data_Large_OA.xlsx

   • Example fields: CustomerID, LastLoginDate, ServiceUsage, etc.

5. Churn Labels – Customer_Churn_Data_Large_CH.xlsx

   • Example fields: CustomerID, ChurnStatus (0 = Not churned, 1 = Churned).

Place these files in a suitable folder on your machine and update the file paths in the notebook if needed (see the Getting Started section).

---

3. Repository Structure

A suggested structure for your GitHub repo:

Customer_Churn_Analysis/
├─ Customer_Churn_Analysis.ipynb        # Main analysis notebook (original or updated)
├─ Customer_Churn_Analysis_updated.ipynb# Updated notebook with image saving
├─ images/                              # Auto-saved plots from the notebook
│  ├─ age_distribution.png
│  ├─ age_by_income_level.png
│  ├─ customers_by_income_level.png
│  ├─ gender_distribution.png
│  ├─ amount_spent_distribution.png
│  ├─ spending_by_product_category.png
│  ├─ age_vs_amount_spent_by_gender.png
│  ├─ age_amount_correlation_matrix.png
│  ├─ amount_spent_by_marital_status.png
│  ├─ age_distribution_by_churn_status.png
│  ├─ churn_by_gender.png
│  ├─ income_level_by_churn_status.png
│  ├─ amount_spent_by_churn_status.png
│  ├─ transaction_frequency_by_churn_status.png
│  ├─ demographics_correlation_matrix.png
│  └─ amount_spent_boxplot_outliers.png
└─ README.md

---

4. Getting Started

4.1. Requirements

Main Python libraries used:

• pandas
• numpy
• matplotlib
• seaborn
• scikit-learn
• jupyter / notebook

Install them (for example):

pip install pandas numpy matplotlib seaborn scikit-learn notebook

4.2. Project Path & Data Files

On the local machine where this notebook was developed, the project root is:

C:\Projects\Customer_Churn_Analysis

The notebook expects the Excel files to be reachable from your current working directory. For portability, you can place all Excel files in a data/ folder in your project and update the paths in the notebook. Example:

# Example (edit paths as needed)
df_demo = pd.read_excel(r'data/Customer_Churn_Data_Large_CD.xlsx')
df_trans = pd.read_excel(r'data/Customer_Churn_Data_Large_TH.xlsx')
df_cs   = pd.read_excel(r'data/Customer_Churn_Data_Large_CS.xlsx')
df_oa   = pd.read_excel(r'data/Customer_Churn_Data_Large_OA.xlsx')
df_ch   = pd.read_excel(r'data/Customer_Churn_Data_Large_CH.xlsx')

4.3. Running the Notebook

From the project root (C:\Projects\Customer_Churn_Analysis or your chosen folder):

jupyter notebook

Open Customer_Churn_Analysis_updated.ipynb and run the cells in order.
The notebook will automatically create an images/ folder (if it doesn’t exist) and save plots into it.

---

5. Exploratory Data Analysis (EDA)

The EDA section covers:

• Demographic distributions

  • Age distribution of customers.
  • Income level and gender distributions.

• Transaction behaviour

  • Amount spent distribution.
  • Spending by product category.

• Relationships and correlations

  • Age vs. Amount Spent (coloured by gender).
  • Correlation between key numerical variables.

• Churn-related behaviour

  • Age vs churn.
  • Gender vs churn.
  • Income level vs churn.
  • Amount spent vs churn.
  • Transaction frequency vs churn.

5.1. Example Visualisations (from images/)

Once you run the updated notebook, several plots are saved to the images/ folder. You can embed some key ones in your GitHub README like this:

---

Age by Income Level

---

Age Distribution

---

Age distribution by Churn

---

Age and Income Correlation

---

Amount Spent by Marital Status

---

Customers by Income Level

---

Demographics Correlation Matrix

---

Gender Distribution

---

Income level by Churn Status

---

Spending by Product Ccategory

---

Transaction Frequency by Churn Status

---

You can pick whichever plots you feel are most “unique” or insightful for your repo.

---

6. Feature Engineering & Preprocessing

Key steps in the notebook:

1. Merging datasets
   All datasets are merged on CustomerID:

   df = df_demo.merge(df_trans, on='CustomerID', how='left')
   df = df.merge(df_cs, on='CustomerID', how='left')
   df = df.merge(df_oa, on='CustomerID', how='left')
   df = df.merge(df_ch, on='CustomerID', how='left')

2. Handling missing values

   • Numerical columns: imputed using the mean.
   • Remaining rows with critical missing values: dropped if necessary.

   df = df.fillna(df.mean())  # numeric imputation
   df = df.dropna()           # drop rows with remaining NaNs
   final_df = df

   (Earlier in the notebook, there is also discussion of imputing categorical values like IncomeLevel with the most frequent category.)

3. Encoding categorical variables
   Categorical features across the different dataframes are encoded using pd.get_dummies, for example:

   df_oa = pd.get_dummies(df_oa, columns=['ServiceUsage'], drop_first=True)

4. Scaling / transformation
   For transaction data, the notebook demonstrates both:

   • Standardisation (Z-score) with StandardScaler
   • Normalisation (Min-Max) with MinMaxScaler
   • A log-transformed version of AmountSpent (to reduce skewness).

   These transformations are used to stabilise variance and make model training more robust.

---

7. Model Development

The model used is a Random Forest Classifier.

7.1. Feature / Target Split

X = df.drop('ChurnStatus', axis=1)
y = df['ChurnStatus']

7.2. Train–Test Split

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

7.3. Model Training

from sklearn.ensemble import RandomForestClassifier

model = RandomForestClassifier(random_state=42)
model.fit(X_train, y_train)

7.4. Evaluation

The notebook evaluates both:

• Test data performance

  from sklearn.metrics import accuracy_score, classification_report
  
  y_pred = model.predict(X_test)
  accuracy = accuracy_score(y_test, y_pred)
  print(f"test_Accuracy: {accuracy}")
  print("Classification Report:")
  print(classification_report(y_test, y_pred))

• Training data performance

  y_train_pred = model.predict(X_train)
  train_accuracy = accuracy_score(y_train, y_train_pred)
  print(f"Training Accuracy: {train_accuracy}")
  print("Training Classification Report:")
  print(classification_report(y_train, y_train_pred))

---

8. Model Performance Summary

From the report section in the notebook:

• Model: Random Forest Classifier

• Example training metrics (illustrative from your summary):

  • Training accuracy around 0.9875.
  • For the churn class (1):
    • High precision (~1.00).
    • Strong recall (around 0.93).
    • F1-score around 0.97.

• The model shows:

  • Very high overall accuracy.
  • Excellent ability to detect churners (class 1) with strong F1-score.
  • Minimal false positives, meaning retention efforts are more targeted.

These results are discussed in the notebook in terms of:

• Strengths of the model.
• Slight overfitting risk (training vs test comparison).
• Business implications for retention strategies.

---

9. How Images Are Saved (Important for GitHub)

The updated notebook defines an IMAGES_DIR at the top:

import os

# Create images directory relative to the project root
IMAGES_DIR = os.path.join(os.getcwd(), "images")
os.makedirs(IMAGES_DIR, exist_ok=True)

For each key plot, code like this has been added before plt.show():

plt.tight_layout()
plt.savefig(os.path.join(IMAGES_DIR, "age_distribution.png"), bbox_inches="tight")
plt.show()

This means:

• When you run the notebook from your project root, all images automatically appear in the images/ folder.
• You can directly reference them in this README.md.

---

10. Future Improvements

Possible extensions:

• Hyperparameter tuning for the Random Forest (e.g., GridSearchCV).
• Feature importance visualisation and SHAP analysis.
• Handling class imbalance with techniques such as SMOTE or class weights.
• Building alternative models (XGBoost, LightGBM, Logistic Regression) for comparison.
• Packaging code into reusable Python modules and adding unit tests.

---

11. Acknowledgements

• Data: Synthetic/organisational customer churn dataset (Customer_Churn_Data_Large_*.xlsx series).
• Author: Ahmed Kansulum Mohammed (Hexatech Solutions / Tech Dynamics Integrated).