Customer Churn Prediction - Project Details

Overview

This project develops a machine learning model to predict customer churn in the telecom industry. Using historical data—demographics, service usage, billing, and tenure—I built a logistic regression model to identify customers at risk of leaving. The goal? Help telecom companies retain customers proactively, boosting satisfaction and cutting acquisition costs. Implemented in a Jupyter notebook, it combines data cleaning, visualization, and evaluation to deliver actionable insights.

Churn Distribution: Bar plot showing the ratio of customers who churned vs stayed

Confusion Matrix: Heatmap showing model prediction accuracy

ROC Curve: Evaluates the model's ability to distinguish between churn and non-churn customers

Report: model performance for predicting both churn and no-churn cases

churner — Top 10 customers who are most likely to churn based on the model's predictions

Data Sources

Telco Customer Churn Dataset: Sourced from kaggle (https://www.kaggle.com/datasets/yeanzc/telco-customer-churn-ibm-dataset) telco-customer-churn.csv

Technologies Used

Environment: Jupyter Notebook, Python
Libraries:
- Data Handling: Pandas, NumPy
- Visualization: Matplotlib, Seaborn
- Machine Learning: Scikit-learn

Key Features

Cleans data by removing duplicates, handling missing values, and encoding categories (e.g., one-hot encoding)
Visualizes churn patterns (bar plots), model accuracy (confusion matrix), and predictive power (ROC curve)
Trains a logistic regression model for binary classification (churn/no churn)
Outputs churn probabilities and identifies high-risk customers

Analysis Process

Data Prep: Load telco-customer-churn.csv, drop customerID, encode categorical variables with pd.get_dummies(drop_first=True)
Exploration: Plot churn distribution (sns.countplot)
Modeling: Split data (80/20), train logistic regression (LogisticRegression), predict churn
Evaluation: Assess with accuracy (82%), AUC (0.86), confusion matrix, and ROC curve

Model Performance

Accuracy: 82% (correctly predicts churn/no churn)
AUC: 0.86 (strong ability to distinguish churners)
Metrics: Precision, recall, F1-score via classification_report; visualized with confusion matrix and ROC curve

Challenges Overcome

Handled high-dimensional data (6,560 columns post-encoding) with efficient preprocessing
Balanced interpretability and performance using logistic regression over complex models
Managed categorical data complexity with one-hot encoding and drop_first=True to avoid multicollinearity

Benefits

Proactive Retention: Pinpoints at-risk customers for targeted interventions
Cost Efficiency: Reduces churn to lower acquisition costs
Customer Satisfaction: Enables tailored retention strategies

Future Improvements

Test advanced models (e.g., Random Forest, XGBoost) for higher accuracy
Integrate real-time CRM data for dynamic predictions
Explore feature engineering (e.g., interaction terms) to boost performance

Try It

Clone the repo and run the notebook yourself! See setup instructions on GitHub.

Visit GitHub Repo