Telco Customer Churn Prediction

Python
Machine Learning
Predictive Modeling
Web App
An end-to-end machine learning project that involves cleaning, training, and deploying an interactive web application to predict customer churn on the classic Telco dataset.
Author

Christopher Hynes

Published

July 14, 2025

Live Application | View the Notebook | GitHub Repository

Project Overview

This project demonstrates a complete machine learning workflow for predicting customer churn for TElco, a telecom company. It involves data cleaning, model training, feature importance analysis, and the deployment of an interactive web application using Streamlit.

By identifying at-risk customers, the business can proactively offer incentives or support to retain them, which is often more cost-effective than acquiring new customers.

Key Features

  • Interactive Prediction: Use sliders and dropdowns to input customer data and receive a real-time churn prediction.
  • Clear Results: The app displays a clear, understandable prediction of whether a customer is likely to churn.
  • Prediction Confidence: Shows the model’s confidence score (probability) for each prediction.
  • Optimized Model: Powered by a pruned logistic regression model that has been optimized through feature selection for efficiency and interpretability.

Methodology

The project follows a standard data science workflow:

a. Data Cleaning and Preprocessing

  • The TotalCharges column was converted to a numeric type, and rows with missing values were dropped.
  • The non-predictive customerID column was removed.
  • The target variable Churn was converted to a binary format (1/0).

b. Model Training and Feature Selection

  • A Logistic Regression model was chosen for its effectiveness and high interpretability.
  • A ColumnTransformer pipeline was used to automatically apply StandardScaler to numerical features and OneHotEncoder to categorical features.
  • Feature Importance Analysis: The model’s coefficients were analyzed, revealing that gender, PhoneService, and Partner had a negligible impact on predictive power.
  • Model Pruning: A second, simpler model was trained with these unimportant features removed. This pruned model achieved the same performance (F1-score of 0.61) and was selected for the final application.

Tools and Libraries

  • Data Analysis & Modeling: Pandas, Scikit-learn, Jupyter
  • Web Application: Streamlit
  • Plotting: Matplotlib, Seaborn