Telecom Churn Prediction Project

Overview

This project analyzes customer data from a telecommunications company to predict customer churn using machine learning. The goal is to identify high-risk customers and enable targeted retention strategies.

The workflow includes:

• Data exploration and preprocessing
• Feature engineering
• Model training and evaluation
• Customer segmentation
• Model interpretation with SHAP
• Deployment-ready model persistence

---

Dataset

Source

Kaggle: Telco Customer Churn

Key Features

Feature	Description
customerID	Unique identifier
gender, SeniorCitizen, Partner, Dependents	Demographic information
tenure	Months with the company
PhoneService, InternetService, OnlineSecurity, etc.	Service usage
Contract, PaperlessBilling, PaymentMethod	Billing details
MonthlyCharges, TotalCharges	Financial metrics
Churn	Target variable (Yes/No)

Dataset Statistics

• Rows: 7,032 customers
• Columns: 21 features
• Churn distribution: 5163 (No), 1869 (Yes)
• Class imbalance: Approximately 26% churn rate

---

Methodology

Exploratory Data Analysis (EDA)

Initial Observations

• Tenure distribution: Median tenure is 29 months
• Monthly charges: Range from $18.25 to $118.75
• Total charges: Skewed distribution with many low values

Key Insights

• Customers with month-to-month contracts have higher churn rates
• Fiber optic internet users show higher churn when lacking support services
• Electronic check payment method correlates with significantly higher churn (45.3%)
• Customers without online security have 41.8% churn rate vs. 14.6% with security

Visualizations

• Churn distribution by categorical variables
• Tenure vs. monthly charges scatter plot
• Correlation matrix of numerical features
• Heatmaps showing churn rates by contract type and payment method

---

Feature Engineering

Derived Features

• NumServices: Count of services used (0-6)
• tenure_group: Categorized as New (<1yr), Mid (1-3yrs), Long (>3yrs)
• ContractNum: 0=Month-to-month, 1=One year, 2=Two year
• PaymentRisk: Risk score for payment methods (Electronic check = highest risk)
• HasFiberAndNoSupport: Flag for high-risk combination of fiber optic + no tech support

Preprocessing Pipeline

• Numerical features: StandardScaler normalization
• Categorical features: OneHotEncoder with drop='first' to avoid multicollinearity
• Missing values: TotalCharges filled with 0 after conversion to numeric

---

Model Development

Evaluation Strategy

• Cross-validation: Stratified K-Fold (5 splits)
• Metrics: ROC AUC, Precision, Recall, F1-score
• Training split: 80% train, 20% test (stratified)
• Random state: 42 for reproducibility

Models Compared

Model	Test AUC	Recall (Churn)	Precision (Churn)	F1 (Churn)
Logistic Regression	0.834	50.8%	65.3%	57.1%
Random Forest	0.821	46.3%	67.3%	54.8%
XGBoost	0.807	67.1%	50.7%	57.8%
LightGBM	0.828	74.1%	50.5%	60.1%
CatBoost	0.830	77.0%	51.5%	61.7%

Champion Model Selection

• Selected: CatBoost Classifier
• Rationale: Highest recall for churn class (77.0%), critical for identifying at-risk customers
• Configuration:
  • scale_pos_weight: 2.75 (accounts for class imbalance)
  • random_state: 42
  • verbose: 0 (suppresses output during training)

---

Model Performance

Final Results

• Test AUC: 0.830
• Accuracy: 75%
• Classification Report:

  precision    recall  f1-score   support
           0       0.90      0.74      0.81      1033
           1       0.52      0.77      0.62       374
  accuracy                           0.75      1407
  macro avg       0.71      0.75      0.71      1407
  weighted avg       0.80      0.75      0.76      1407
  

Interpretation

• Recall of 77%: Model successfully identifies 77% of actual churners
• Precision of 52%: Half of predicted churners actually churn
• F1-score of 62%: Balanced measure of precision and recall
• AUC of 0.83: Strong discriminatory power between churn and non-churn customers

---

Customer Segmentation

Clustering Approach

• Algorithm: K-Means clustering
• Features: tenure, MonthlyCharges, NumServices, ContractNum, PaymentRisk, SeniorCitizen, Partner
• Optimal clusters: k=4 determined via elbow method

Segment Profiles

Segment	Characteristics	Churn Risk
0	Long-term, high spend, multiple services	Low
1	Low spend, few services, stable contracts	Medium
2	Seniors, medium spend, limited services	High
3	New customers, high churn risk	Critical

---

Model Interpretability

SHAP Analysis

• Purpose: Understand feature contributions to predictions
• Method: SHAP values on a sample of 100 customers
• Key Findings:
  • Negative impact: tenure, contract length, payment method risk
  • Positive impact: fiber optic service, lack of support services
  • Most influential features: tenure, contract type, payment method

---

Implementation

File Structure

TelecomChurnPrediction/
├── models/
│   ├── champion_model.pkl
│   ├── preprocessor.pkl
│   ├── kmeans.pkl
│   ├── scaler.pkl
│   ├── elbow_plot.png
│   └── shap_summary.png
├── app.py
├── streamlit_app.py
├── requirements.txt
├── Telco_Customer_Churn.ipynb
└── README.md

Dependencies

pandas
numpy
scikit-learn
catboost
seaborn
matplotlib
shap
joblib
streamlit
fastapi
uvicorn

Model Persistence

• Format: joblib pickle files
• Files saved:
  • champion_model.pkl: Trained CatBoost model
  • preprocessor.pkl: Feature preprocessing pipeline
  • kmeans.pkl: Customer segmentation model
  • scaler.pkl: StandardScaler for clustering
• Version compatibility: Models trained with scikit-learn 1.6.1

---