I embarked on a data‑science learning project to predict churn and segment customers using the public Telco Customer Churn dataset. The workflow covered data cleaning, exploratory analysis, hypothesis testing, model training, hyperparameter tuning, SHAP explainability, K‑Means clustering, and risk‑tier categorisation.
The dataset contains 7,043 customer records.
After hyperparameter tuning, the XGBoost model achieved a ROC‑AUC of 0.85 on 5‑fold cross‑validation.
The newer, lower‑charge segment showed a churn rate of 31.39 %, while the long‑tenure, higher‑charge segment churned at 16.91 %.
I chose XGBoost because it outperformed Logistic Regression and Random Forest, and I used SHAP to surface the most influential features for business stakeholders. K‑Means clustering on tenure, monthly charges, and total charges produced interpretable segments that map to low, medium, and high churn‑risk tiers for targeted retention actions. My LinkedIn page documents this project among others.