📊 Customer Churn Prediction — Machine Learning Project

This project focuses on understanding customer churn for a DTH/E-Commerce service provider and building predictive models to identify customers who are likely to leave. By analyzing customer behavior, service usage patterns, revenue metrics, complaints, and demographic factors, the project provides a data-driven framework for churn reduction strategies.

---

🚀 Project Overview

Customer churn is a major challenge for subscription-based industries. Retaining customers is significantly cheaper than acquiring new ones — making churn prediction a critical business need.

Using a dataset of 11,260 customer accounts and 19 attributes, this project performs:

• Exploratory Data Analysis (EDA)
• Data cleaning & feature engineering
• Handling of imbalanced data using SMOTE
• Model building using multiple ML algorithms
• Feature importance analysis
• Business recommendations based on insights

---

📂 Dataset Summary

• Total Records: 11,260
• Target Variable: Churn (1 = churned, 0 = not churned)
• Features Include:
  • Tenure
  • City tier
  • Revenue per month
  • Payment mode
  • Complaint history
  • Account segment
  • Customer service interactions
  • Cashback, coupon usage
  • Login device preference

---

🔍 Exploratory Data Analysis (EDA)

Key steps included:

• Treatment of special characters and incorrect data types
• Handling missing values (median/mode imputation)
• Outlier detection
• Univariate, bivariate & multivariate analysis
• Pairplots & correlation heatmaps
• Identification of top churn drivers

Important observations:

• Low-tenure customers churn the most
• Customers with past complaints have higher churn probability
• “Regular Plus” accounts show maximum churn
• “Super” accounts show highest retention
• Churn dataset is highly imbalanced (84%:16%)

---

⚙️ Models Used

Four classification algorithms were evaluated:

1. Random Forest Classifier
2. Logistic Regression
3. AdaBoost Classifier
4. XGBoost Classifier (🏆 Best Model)

---

🏆 Best Performing Model: XGBoost

XGBoost delivered the highest performance across accuracy, recall, and precision.

Why XGBoost performed best:

• Handles non-linear relationships
• Robust to imbalanced datasets (with proper tuning)
• Strong feature importance insights
• High predictive power

Top Predictive Features:

• Tenure
• Complain_l12m
• Days_since_CC_connect
• Account_segment

---

📈 Model Evaluation

Model	Notes
Random Forest	Good recall but biased towards non-churn class
Logistic Regression	Lower accuracy; weaker feature separation
AdaBoost	Weak performance on imbalanced data
XGBoost	Highest accuracy and best overall performance

---

🧠 Business Insights

Based on modeling and data analysis:

• Offer long-tenure plans to reduce churn among new customers
• Strengthen customer service responsiveness — complaints strongly correlate with churn
• Promote Super-tier plans given their stronger retention
• Targeted retention offers for Regular Plus accounts
• Run periodic surveys for marital status correlations with churn behavior

---

📌 Files in This Repository

• notebooks/ — Jupyter notebooks (EDA, cleaning, model training)
• data/ — (If allowed) processed datasets
• models/ — Trained model files (optional)
• reports/ — PDF report, presentation, visual summaries
• README.md — Project documentation

---

🛠️ Tech Stack

• Python
• Pandas, NumPy
• Matplotlib, Seaborn
• Scikit-learn
• XGBoost
• Imbalanced-learn (SMOTE)

⭐ If you find this useful, please give the repository a star!