Predicting Customer Churn.

A machine learning solution designed to predict customer churn and help the marketing team retain at-risk users through personalized offers.

  • CLIENT:

    Interconnect

  • DATE:

    June 03, 2025

  • PROJECT:

    GitHub

Exploration and Processing

We worked with a dataset of 7,000 users, containing details such as contract type, subscribed services, payment method, billing frequency, and demographic attributes (e.g., couple, dependents, senior citizen). Irrelevant columns including CustomerID, StartDate, EndDate, and Gender were removed to keep only meaningful features.

During preprocessing, we identified rows with missing or whitespace values in TotalCharges and removed them before converting the column to numeric format.

In feature engineering, categorical variables such as Type, PaymentMethod, and InternetService were transformed using One-Hot Encoding, while numerical variables (MonthlyCharges, TotalCharges) were standardized with a StandardScaler. Binary fields were converted from “Yes/No” to 1/0 for consistency.

Because the target variable (churn) was imbalanced — with 73% active users and ~26% churned users — we applied a combination of Random UnderSampling and SMOTE to create a balanced training set.

Model Training and Findings

Three machine learning models were trained and evaluated using pipelines with ColumnTransformer:

  • Logistic Regression (baseline)
  • Random Forest
  • XGBoost

A GridSearchCV was used for hyperparameter tuning, and cross-validation ensured robust performance estimates.

Model performance was assessed using Accuracy and ROC AUC on both training and test sets:

  • Logistic Regression → 0.84 / 0.84
  • Random Forest → 0.86 / 0.84
  • XGBoost → 0.87 / 0.85

All models demonstrated consistent performance between training and testing sets, indicating no overfitting.

The final model selected was XGBoost, achieving the highest ROC AUC score and proving to be the most effective at predicting customer churn with precision.

Distribution of churn across key categorical variables. This analysis helped identify which features show meaningful differences between churned and active users. For example, gender showed nearly identical distributions, suggesting it has little predictive value.
Distribution of churn across key categorical variables. This analysis helped identify which features show meaningful differences between churned and active users. For example, gender showed nearly identical distributions, suggesting it has little predictive value.
Distribution of target classes before and after applying RandomUnderSampler and SMOTE. This visualization highlights how the dataset was balanced to ensure equal representation of churned and active users, improving model training stability.
Distribution of target classes before and after applying RandomUnderSampler and SMOTE. This visualization highlights how the dataset was balanced to ensure equal representation of churned and active users, improving model training stability.
ROC AUC curve for Logistic Regression. The model shows solid baseline performance with balanced sensitivity and specificity, achieving an AUC of 0.84.
ROC AUC curve for Logistic Regression. The model shows solid baseline performance with balanced sensitivity and specificity, achieving an AUC of 0.84.
ROC AUC curve for Random Forest. This ensemble model improves stability and slightly enhances predictive performance compared to Logistic Regression, reaching an AUC of 0.84.
ROC AUC curve for Random Forest. This ensemble model improves stability and slightly enhances predictive performance compared to Logistic Regression, reaching an AUC of 0.84.
ROC AUC curve for XGBoost. The model delivers the strongest performance among all tested algorithms, achieving an AUC of 0.85 and becoming the final choice for churn prediction.
ROC AUC curve for XGBoost. The model delivers the strongest performance among all tested algorithms, achieving an AUC of 0.85 and becoming the final choice for churn prediction.