Project

Customer Churn Classifier — Telco

End-to-end XGBoost classifier on the public IBM Telco churn dataset — scikit-learn Pipeline + ColumnTransformer, SHAP explainability, a model card, and an interactive Streamlit demo. ROC-AUC 0.842.

title: Customer Churn Classifier — Telco blurb: End-to-end XGBoost classifier on the public IBM Telco churn dataset — scikit-learn Pipeline + ColumnTransformer, SHAP explainability, a model card, and an interactive Streamlit demo. ROC-AUC 0.842. personas: [data] stack:

• Python
• scikit-learn
• XGBoost
• SHAP
• Streamlit
• pandas
• numpy
• matplotlib
• seaborn links: github: https://github.com/SyrineLarbi/churn-classifier demo: https://churn-classifier-nk8qxgpsbeh8akhk3luqwu.streamlit.app/

What it is

A supervised classifier that flags which telecom customers are likely to churn, built on the public IBM Telco Customer Churn dataset. The whole path — preprocessing, training, evaluation, explainability, and a live demo — is reproducible from a single git clone.

Highlights

• Performance — ROC-AUC 0.842, PR-AUC 0.652, accuracy 0.81 on a 20% stratified holdout. Churn recall of 0.55, tuned to catch at-risk customers rather than chase headline accuracy on an imbalanced target.
• Explainability — SHAP summary, bar, and waterfall plots turn the XGBoost model into a story: tenure, month-to-month contracts, fiber-optic service, monthly charges, and electronic-check payment are the top churn drivers by mean absolute SHAP value.
• Live demo — a Streamlit app with slider controls lets anyone score a hypothetical customer and see the prediction change in real time.
• Documented — a model card records the dataset, metrics, intended use, and limitations, so the numbers are honest about what the model can and can't do.

What it took

• Preprocessing — one scikit-learn Pipeline + ColumnTransformer handling numeric scaling and categorical encoding, so the exact same transforms run at training and inference time — no train/serve skew.
• Training & evaluation — XGBoost fit in ~3 seconds, evaluated with ROC and PR curves plus a confusion matrix rather than accuracy alone, because churn is an imbalanced problem where accuracy flatters.
• Reproducibility — pinned dependencies, a fast training script, and saved figures so the README and model card embed real plots. Clone, install, run end-to-end.

Stack at a glance

scikit-learn (Pipeline + ColumnTransformer) · xgboost · shap · streamlit · pandas / numpy · matplotlib / seaborn — the smallest set that covers preprocessing, modeling, explainability, and a shippable demo.