Building a Customer Churn Prediction Model with Logistic Regression

This week, I tackled Module 3: Machine Learning for Classification as part of the ML Zoomcamp 2026. The objective was to create a model that could assist a telecom company in identifying customers at risk of leaving. This hands-on project allowed me to link various machine learning concepts to a real-world scenario.

To begin, I delved into the Telco customer churn dataset, analyzing the target variable, inspecting missing values, and observing the distribution of different features. Following this, I investigated the relationship between features and customer churn using several methods: risk ratio and risk difference for categorical variables, mutual information to gauge the informativeness of categorical features about the target, and correlation coefficients for numerical features.

A notable finding was that contract type, among the categorical features examined, possessed the highest mutual information. To make machine learning models work with numerical inputs, I utilized Scikit-Learn's DictVectorizer to carry out one-hot encoding on categorical features, resulting in a feature matrix with 45 columns.

Next, I trained a logistic regression model using Scikit-Learn, which achieved around 80% accuracy on the validation set. Additionally, I explored how to interpret the model's weights, trained a simplified model using a subset of features, and worked on calculating predictions manually.

The most significant takeaway from this project was comprehending the interplay between the various stages of a machine learning project. Feature analysis aids in understanding the data, encoding prepares it for the model, logistic regression leverages these features to estimate the probability of customer churn, and evaluation helps gauge the model's performance.

This classification workflow can be applied to other problems like spam detection, loan default prediction, and lead scoring. In the coming days, I'll progress to Module 4: Evaluation Metrics for Classification, where I'll delve deeper into measuring model performance and understanding that accuracy alone may not always provide a complete picture. I'm committed to learning these concepts one by one, emphasizing the rationale behind the code rather than merely getting it to function.