Customer Churn Prediction using Machine Learning

Customer Churn Prediction using Machine Learning is a business-analytics final year project that flags which customers are likely to cancel, why they are leaving, and what retention actions could change the outcome. Built on telecom-style customer records (tenure, monthly charges, contract type, support tickets, subscribed services), an XGBoost classifier learns churn patterns while class-imbalance techniques keep the minority churn class from being ignored. Everything is served through a Streamlit dashboard branded ChurnGuard: churn-risk segments, SHAP-ranked churn drivers, revenue-at-risk summaries and a what-if simulator that shows how a retention offer moves an individual customer's risk score. Because churn analytics maps directly to real industry roles, this is a project recruiters — and viva examiners — instantly understand.

This is a built-to-order project: the code, trained models, report, PPT and viva Q&A are prepared and quality-checked for your order.

1. Telecom-style customer records are loaded and cleaned: contract types encoded, charges scaled, tenure binned, missing values imputed.
2. The data is split with stratification so the churn minority keeps its share in train and test sets.
3. An XGBoost classifier is trained with scale_pos_weight set from the class ratio, then tuned (max_depth, learning_rate, min_child_weight) via cross-validation.
4. The decision threshold is tuned on the validation set to favour recall on churners — missing a churner costs more than a false alarm.
5. SHAP values explain every prediction; the ChurnGuard dashboard turns scores, segments, drivers and what-if simulations into an interactive retention tool.

• Per-customer churn probability scoring with adjustable decision threshold
• Risk segmentation (high / medium / low) with customer counts and revenue-at-risk rollups
• SHAP-ranked churn drivers — e.g. month-to-month contract, short tenure, high monthly charges, frequent support tickets
• What-if retention simulator: change contract type, tenure or support-call count and watch the risk score move
• Class-imbalance handling: scale_pos_weight, stratified splits, optional SMOTE — compared explicitly in the notebook
• Proper classification metrics: precision, recall, F1, ROC-AUC and PR curves, with threshold tuning aimed at churn-class recall
• Streamlit dashboard (ChurnGuard) with filters by segment, contract type and tenure band
• Retention-action mapping: each top driver links to a suggested business action (contract upgrade offer, proactive support call)

• Full source code (training notebooks + Streamlit dashboard, well-commented)
• Trained model files (.pkl) — the dashboard runs without retraining
• Project report PDF (synopsis, literature survey, methodology, results, conclusion, business interpretation)
• PPT presentation for the final review
• Viva Q&A preparation document (imbalance handling, XGBoost internals, SHAP, precision-recall trade-offs)
• Setup guide (environment, dependencies, running the dashboard locally)

FAQs

1. Why not just report accuracy for this churn prediction project? With ~70%+ non-churners, a model predicting "nobody churns" scores 70% accuracy while being useless. The project optimizes and reports precision, recall, F1 and ROC-AUC instead — and the viva Q&A arms you to explain exactly why accuracy misleads on imbalanced data.
2. How is class imbalance handled — SMOTE or scale_pos_weight? Both are implemented and compared: scale_pos_weight (cheap, no synthetic data) versus SMOTE oversampling. You can present the trade-off table in your review — examiners love this question.
3. What does SHAP add beyond feature importance? Built-in importance says which features matter globally; SHAP explains individual predictions — why this customer is high-risk — which powers the what-if simulator and the retention-action mapping.
4. How is the decision threshold chosen? By scanning thresholds on the validation set and picking the point that maximizes churn-class recall subject to an acceptable precision floor, since a missed churner costs more than an unnecessary retention call.
5. Can I run the dashboard offline for my review? Yes — trained models ship as .pkl files and Streamlit runs locally, so the full demo works without internet.
6. What extensions impress examiners? Cohort analysis (churn by tenure band over time), a cost-based threshold optimizer using real retention-offer costs, or survival analysis for when a customer will churn. Ask us to scope any of these.

• The model finds correlational churn drivers, not proven causes — a retention offer's true effect needs an A/B test, which the what-if simulator approximates, not guarantees.
• Demo data is telecom-style; results on a real company's data would differ in feature mix and churn base rate.
• Churn behaviour drifts (new plans, competitors), so the model needs periodic retraining on fresh data.
• Revenue-at-risk figures are illustrative estimates from charges in the dataset, not audited financials.