Customer Churn Prediction and Segmentation

Overview

This project predicts customer churn and segments customers for targeted retention strategies. It provides:

• A training and evaluation pipeline for churn prediction (RandomForest, XGBoost, LightGBM).
• K-Means segmentation including churn probability as a clustering feature.
• An interactive Streamlit dashboard for exploration and suggested retention actions.

Business Impact

This project provides a dual approach to tackling customer churn. The predictive models enable proactive identification of individual customers at risk of churning, allowing for timely intervention. Simultaneously, the clustering analysis offers a strategic understanding of why different customer segments churn, facilitating the development and implementation of highly tailored and cost-effective retention strategies. By understanding both who is likely to churn and which groups share common churn-driving characteristics, the business can optimize resource allocation and improve customer satisfaction and loyalty, ultimately reducing overall churn.

Files & Key Components

• app.py — Streamlit dashboard. Key functions include:
  • app.load_data — Loads Data/customer_data_with_clusters.csv.
  • app.cluster_summary — Generates per-cluster aggregated metrics and churn lift.
  • app.retention_priority_score — Heuristic priority score.
  • app.strategy_for_cluster — Suggested actions for each cluster.
• [Customer Churn Prediction and Segmentation.ipynb](Customer Churn Prediction and Segmentation.ipynb) — Notebook with data cleaning, feature engineering, modeling, and clustering steps. It creates:
  • Data/customer_data_with_clusters.csv (used by the dashboard)
  • Saved models in Model_Artifacts/ (e.g., LightGBM and KMeans models).
• Data:
  • Data/WA_Fn-UseC_-Telco-Customer-Churn.csv — Original dataset (Telco Customer Churn).
  • Data/customer_data_with_clusters.csv — Processed dataset with Churn_Probability and Cluster columns.
• Model_Artifacts/ — Model files (e.g., LightGBM_churn_model.pkl, KMeans_churn_clusters.pkl).

How to run

1. Install dependencies:

pip install -r requirements.txt

2. If you want to create the processed dataset and models, open and run the notebook:

• [Customer Churn Prediction and Segmentation.ipynb](Customer Churn Prediction and Segmentation.ipynb)

3. Launch the dashboard (it expects Data/customer_data_with_clusters.csv to be present):

streamlit run app.py

• Local link: http://localhost:8501
• Deployed link: <https://telco-churn-strategy.streamlit.app/>

Dashboard features

The Streamlit app (app.py) provides:

• Global KPIs: Selected customers, average churn probability, avg monthly revenue, avg tenure.
• Cluster summary and churn lift computed with app.cluster_summary.
• Priority scoring per cluster using: $$ score = avg_churn \times avg_monthly \times \log(1 + cluster_size) $$ (Implemented by app.retention_priority_score)
• Suggested retention actions per cluster (via app.strategy_for_cluster).
• Visual insights: cluster bar chart + churn line, MonthlyCharges vs Churn scatter, contract vs payment heatmaps, feature differences vs global mean, churn by tenure buckets.
• A top-customers list (risk x value) and CSV download for the selected filters.

Model comparison

All metrics below are from the notebook experiments in the repository. Use them as a reference — metric values are based on the notebook's runs.

Model	Test Accuracy	Train Accuracy	Recall (Churn=1)	AUC Score
RandomForest	0.7270788913	0.7500444444	0.81	0.8303044971
XGBoost	0.7007818053	0.7239111111	0.83	0.8249439616
LightGBM	0.7277896233	0.7575111111	0.82	0.8364363698

Recommendation:

• LightGBM shows the best trade-off (highest AUC and highest test accuracy) and is selected as the primary model in the notebook.
• XGBoost had the best recall (slightly higher recall), which may be preferred if the business objective prioritizes identifying as many churners as possible.

Key findings

• LightGBM performs best overall (AUC ≈ 0.836).
• High-risk clusters (per notebook) are often characterized by:
  • Month-to-month contracts
  • Electronic check payment method
  • Fiber optic internet service
  • Lower tenure and total charges
• Clustering (KMeans, $K=4$) helps prioritize groups for retention campaigns.
• Priority score weights churn probability, average monthly ARPU, and cluster size logarithmically; it helps identify high-impact clusters to focus on.

Summary:

• Clustering identified distinct customer segments: The project successfully segmented customers into 4 clusters. Clusters 0 and 3 were identified as high-churn risk groups, exhibiting average churn probabilities of approximately 0.66 and 0.58, respectively.
• High-churn clusters share specific characteristics: Customers in high-risk Clusters 0 and 3 are predominantly on month-to-month contracts, use electronic check payment methods, and often have fiber optic internet service, coupled with lower average tenure and total charges.
• LightGBM is the top-performing churn prediction model:
  • AUC Score: LightGBM achieved the highest AUC score of 0.836, indicating its superior ability to distinguish between churners and non-churners.
  • Recall (Churn=1): LightGBM demonstrated a strong Recall of 0.82 for the churn class, closely followed by XGBoost (0.83) and RandomForest (0.81), indicating its effectiveness in identifying actual churners.
  • Accuracy: LightGBM also showed the highest Test Accuracy (0.728) and Train Accuracy (0.758), suggesting good generalization.
• Model performance comparison:
  • XGBoost excelled in identifying actual churners with the highest Recall (0.83).
  • RandomForest showed a balanced performance with a Recall of 0.81 and an AUC of 0.830.

Reproducibility notes

• The notebook writes out Data/customer_data_with_clusters.csv and saves model artifacts to Model_Artifacts/. If you want an identical dashboard experience, ensure these outputs are present or re-run the notebook to regenerate them.
• The dashboard (app.py) reads Data/customer_data_with_clusters.csv via app.load_data.

License & Credits

• Dataset: Telco Customer Churn dataset (Kaggle).
• This project is developed by Yousef Mohtady.