Telecommunication Customer Churn Prediction

Final Year Project — Minors in AI/ML Programme, DJ Sanghvi College of Engineering

A complete machine learning pipeline for predicting customer churn in the telecommunications industry. The project explores data analysis, feature engineering, model comparison, hyperparameter tuning, and neural network approaches on a real-world dataset of 7,043 customers.

---

Problem Statement

Customer churn — when subscribers cancel their service — is one of the most expensive problems in telecom. Acquiring new customers costs 5–25x more than retaining existing ones. This project builds predictive models to identify at-risk customers before they leave, enabling targeted retention strategies.

---

Dataset

Property	Value
Total Customers	7,043
Features	21 (demographics, services, account info)
Target	Churn (Yes/No)
Class Split	~73.5% No / ~26.5% Yes

Feature categories:

• Demographics: Gender, Senior Citizen, Partner, Dependents
• Services: Phone, Internet (DSL/Fiber), Online Security, Streaming TV/Movies, Tech Support
• Account: Contract type, Payment method, Tenure, Monthly/Total Charges

---

Methodology

1. Exploratory Data Analysis

• Class imbalance analysis (26.5% churners)
• Boxplots revealing churners have lower tenure (10 months median) and higher monthly charges ($80 median)
• Feature-level countplots identifying high-churn segments:
  • Month-to-month contracts
  • Electronic check payments
  • Fiber optic internet (higher cost, higher churn)
  • No online security or tech support

2. Feature Engineering & Preprocessing

• Label encoding for binary features, one-hot encoding for multi-class
• StandardScaler normalization for numerical features
• Correlation analysis and feature selection
• 80/20 train-test split

3. Baseline Models

• K-Nearest Neighbors (KNN)
• Logistic Regression (L1/L2 regularization)
• Random Forest
• Support Vector Machine (SVM)

4. Hyperparameter Tuning

• GridSearchCV for KNN (optimal k), Logistic Regression (regularization), SVM (C value)
• RandomizedSearchCV for Random Forest (n_estimators, max_depth, max_features, criterion)
• Cross-validation to address train-test split bias

5. Neural Network

• Feed-forward neural network with Keras
• Architecture: Dense layers (1024 → 768 → 512 → 256 → 128 → 1) with Dropout
• ModelCheckpoint for best model selection
• Training history visualization (loss and accuracy curves)

---

Evaluation

Models evaluated using metrics suited for imbalanced classification:

• Accuracy
• ROC AUC Score
• Precision / Recall / F1 Score
• Confusion Matrix
• ROC Curve visualization

Key Insights

• Contract type is the strongest churn predictor — month-to-month customers churn at significantly higher rates
• Tenure is inversely correlated with churn — early intervention is critical
• Fiber optic customers churn more despite (or because of) higher monthly charges
• Customers without online security or tech support add-ons are more likely to churn
• Ensemble and tuned models outperform baseline approaches

---

Tech Stack

• Python: pandas, NumPy, scikit-learn
• Deep Learning: Keras / TensorFlow
• Visualization: Matplotlib, Seaborn (FiveThirtyEight style)
• Tuning: GridSearchCV, RandomizedSearchCV
• Environment: Jupyter Notebook

---

Project Structure

├── Churn Analysis in Telecommunication.ipynb   # Full analysis notebook
├── Customer Churn Prediction.docx              # Project report
├── telecommunication data.csv                  # Dataset
└── README.md

---

How to Run

# Clone the repository
git clone https://github.com/Anuj-G-06/Telecommunication-Churn-Prediction.git
cd Telecommunication-Churn-Prediction

# Install dependencies
pip install pandas numpy matplotlib seaborn scikit-learn keras tensorflow

# Open the notebook
jupyter notebook "Churn Analysis in Telecommunication.ipynb"

---

Author

Anuj Gupta

• Final Year Project, Minors in AI/ML — DJ Sanghvi College of Engineering
• LinkedIn
• GitHub