Customer Churn Analysis

Project Overview

This project focuses on analyzing customer data to predict churn in a telecommunications company. The primary objective is to identify key factors that contribute to customer churn and to build a predictive model that can accurately identify customers who are likely to leave. By understanding the drivers of churn, businesses can implement targeted retention strategies to reduce customer attrition.

---

Table of Contents

• Project Overview
• Objectives
• Data Source
• Methodology
  • 1. Data Loading and Initial Exploration
  • 2. Data Cleaning and Preprocessing
  • 3. Exploratory Data Analysis (EDA)
  • 4. Feature Engineering
  • 5. Model Building
  • 6. Model Evaluation
  • 7. Feature Importance
• Libraries and Tools Used
• Key Findings and Results
• Conclusion
• Future Work
• How to Run
• Contributors

---

Objectives

• To understand the patterns and characteristics of customers who churn versus those who do not.
• To identify the most significant features influencing customer churn.
• To build and evaluate various machine learning models for churn prediction.
• To provide actionable insights for customer retention strategies.
• To analyze the distribution of churn across different customer segments.

---

Data Source

• Dataset : Download
• Description: The dataset was sourced from a telecom provider and contains 7043 customer records with 21 features. It includes customer demographics, account information, services subscribed, and churn status.
• Features: customerID, gender, SeniorCitizen, Partner, Dependents, tenure, PhoneService, MultipleLines, InternetService, OnlineSecurity, OnlineBackup, DeviceProtection, TechSupport, StreamingTV, StreamingMovies, Contract, PaperlessBilling, PaymentMethod, MonthlyCharges, TotalCharges, Churn
• Target Variable: Churn (Yes/No)

---

Methodology

The project follows a standard data science workflow:

1. Data Loading and Initial Exploration

• Loaded the dataset using Pandas from a CSV file.
• Performed initial checks:
  • Data Shape: (7043, 21)
  • Data Types: Mixed numeric and object types.
  • Basic Statistics: Examined numerical features using df.describe().
  • Missing Values Check: Found 11 missing values in TotalCharges.

2. Data Cleaning and Preprocessing

• Missing Values: Handled 11 missing values in TotalCharges by converting to numeric and dropping NA values.
• Data Type Conversion: Converted TotalCharges from object to numeric type.
• Categorical Variable Encoding: Prepared categorical features for analysis (status: not yet encoded in current notebook).
• Feature Scaling: Implemented

3. Exploratory Data Analysis (EDA)

• Churn Distribution: Visualized the target variable, revealing a distribution of 73.4% No Churn and 26.6% Churn.

• Univariate Analysis: Examined distributions of numerical features and counts for categorical features.

• Bivariate Analysis: Created countplots for all categorical features against churn status to observe relationships.

• Correlation Analysis:Implemented

• 

4. Feature Engineering

• Feature Selection: All features were initially considered for analysis.
• New Features: Created a tenure group from the tenure feature which was numerical

5. Model Building

• Data Splitting : use the 80-20 split cross validation
• Models Training : Trained : Logistic Regression , Random Forest , XGBoost, Ada Boost , Gradient Boosting and a simple Neural Network
• Handling Class Imbalance : Used SMOT analysis
• Hyperparameter Tuning : Used Optuna to do hyper-parameter tuning

6. Model Evaluation

• Metrics Used : AUC-ROC

• Model Performance Comparison

• Best Model Performance

7. Feature Importance

We used SHAP and LIME for feature importance analysis

Libraries and Tools Used

• Python >= 3.10
• Pandas: For data manipulation and analysis.
• NumPy: For numerical operations.
• Matplotlib: For basic visualization.
• Seaborn: For advanced statistical visualizations.
• Scikit-learn: Used for classification algorithms (Logistic Regression, Random Forest, Gradient Boosting, AdaBoost, Decision Trees), data preprocessing, model selection, and performance evaluation.
• XGBoost: For optimized gradient boosting models.
• LIME: For local interpretable model explanations.
• SHAP: For model interpretability and feature importance.
• Optuna: For hyperparameter tuning.
• TensorFlow Keras: For building and training neural networks.
• Jupyter Notebook: As the development environment.

---

Conclusion

This project identified key churn drivers and built predictive models to flag high-risk customers. Gradient Boosting outperformed all models based on AUC-ROC and was selected as the final model. SHAP and LIME provided valuable insights into feature importance, supporting strategic business decisions.

Future Work

• Targeted Retention: Offer personalized discounts or upgrades to high-risk customers.
• Proactive Support: Use alerts to flag churn risks and assign dedicated support.
• Customer Engagement: Send usage tips and educational content to boost satisfaction.
• Risk-Based Segmentation: Tailor campaigns by churn risk (high, medium, low).
• Real-Time Scoring: Deploy the model to enable live churn predictions and actions.
• Model Updates: Retrain with new data regularly for improved accuracy.

---

Contributors

• Ronald Nyasha Kanyepi
• GitHub