⏳ Predict Customer Churn

An end-to-end machine learning pipeline for the Kaggle Playground Series Season 6, Episode 3 competition, predicting whether a telecom customer will churn using gradient-boosted decision trees. The project is split across two notebooks: a data processing pipeline and a modelling & evaluation pipeline.

🏆 Competition: Kaggle — Playground Series S6E3
📏 Evaluation Metric: ROC-AUC

---

📌 Project Overview

The goal is to predict binary customer churn (1 = churned, 0 = retained) from a telecom dataset of ~594,000 records. The pipeline focuses on intelligent feature engineering tailored for gradient-boosted trees, followed by a CatBoost baseline model with early stopping and class balancing.

---

📁 Project Structure

├── S6_E3_Predict_Customer_Churn_Data_Processing.ipynb        # Notebook 1: Data pipeline
├── S6_E3_Predict_Customer_Churn__Modeling___Evaluation_.ipynb # Notebook 2: Modelling & submission
├── data/
│   ├── train.csv                  # Raw training data (from Kaggle)
│   ├── test.csv                   # Raw test data (from Kaggle)
│   ├── sample_submission.csv      # Kaggle submission format
│   ├── train_modified.csv         # Processed training data (output of Notebook 1)
│   └── test_modified.csv          # Processed test data (output of Notebook 1)
├── catboost_model.cbm             # Saved CatBoost model
├── predictions.npy                # Saved prediction probabilities
├── meine_submission.csv           # Final Kaggle submission file
└── README.md

---

🗂️ Dataset

Source: Kaggle Playground Series S6E3

Feature	Description
Gender	Customer gender
Tenure	Months as a customer
Partner	Has a partner (Yes/No)
Dependents	Has dependents (Yes/No)
PhoneService	Phone service subscription
MultipleLines	Multiple phone lines
InternetService	Internet service type (DSL / Fiber / None)
OnlineSecurity / OnlineBackup	Add-on services
DeviceProtection / TechSupport	Support services
StreamingTV / StreamingMovies	Streaming subscriptions
Contract	Month-to-month / One year / Two year
PaperlessBilling	Paperless billing flag
PaymentMethod	Payment method used
MonthlyCharges / TotalCharges	Billing amounts
Churn	Target — 1 = churned, 0 = retained

---

🔬 Notebook 1 — Data Processing

File: S6_E3_Predict_Customer_Churn_Data_Processing.ipynb

Key Steps

1. Data Loading

• Downloaded directly from Kaggle API (kaggle competitions download)
• Loaded train.csv and test.csv into pandas DataFrames

2. Feature Engineering — Service Outage Flags

Two new binary features created to distinguish between "not subscribed" vs "service unavailable":

New Feature	Logic	Rationale
InternetOutage	1 if ANY internet service column = 'No internet service'	Differentiates users with no internet access vs no subscription
PhoneServiceOutage	1 if MultipleLines = 'No phone service'	Flags phone service gaps independently

This distinction matters — a user without internet due to a technical issue is fundamentally different from one who simply chose not to subscribe.

3. Encoding

• One-hot encoding on PaymentMethod (4 dummy columns, drop_first=False — all 4 retained as each carries meaningful signal)
• Binary mapping (Yes → 1, No → 0) for 12 binary object columns
• 'No internet service' and 'No phone service' replaced with 'No' after outage flags were created
• Dropped original Gender, InternetService, Contract after transformation

4. Output

• Saved train_modified.csv and test_modified.csv — clean, fully numeric datasets ready for modelling

---

🤖 Notebook 2 — Modelling & Evaluation

File: S6_E3_Predict_Customer_Churn__Modeling___Evaluation_.ipynb

Models Considered

Model	Notes
CatBoost ✅	Baseline — handles categorical features natively, minimal preprocessing required
LightGBM	Planned
XGBoost	Planned
Random Forest	Planned
Naive Bayes	Suitable for binary target prediction
HistGradientBoosting	Hill-climbing approach

Gradient Boosted Decision Trees (GBDTs) are the primary focus — empirically proven to outperform on tabular classification problems of this type.

CatBoost Baseline Configuration

CatBoostClassifier(
    iterations           = 3000,
    early_stopping_rounds = 100,
    learning_rate        = 0.05,
    loss_function        = 'Logloss',
    eval_metric          = 'AUC',
    depth                = 6,
    auto_class_weights   = 'Balanced',   # handles class imbalance
    bagging_temperature  = 0.8,
    random_strength      = 1,
    l2_leaf_reg          = 5
)

Train/test split: Stratified split with test_size=254,655 records, random_state=0

Evaluation

roc_auc = roc_auc_score(y_test, y_proba)

Submission validation:

• Verified id alignment between sample_submission.csv and meine_submission.csv
• Computed MAE between submission and sample as a sanity check

Saved Artefacts

File	Description
catboost_model.cbm	Trained CatBoost model (reload with model.load_model())
predictions.npy	Prediction probabilities array
meine_submission.csv	Final Kaggle submission

---

🚀 How to Run

On Google Colab (recommended — matches original environment)

Notebook 1 — Data Processing:

1. Upload your kaggle.json API key when prompted
2. Run all cells — downloads competition data, processes it, saves train_modified.csv and test_modified.csv

Notebook 2 — Modelling:

1. Data is loaded directly from the GitHub raw URL (no local file needed)
2. Run all cells — trains CatBoost, evaluates ROC-AUC, generates meine_submission.csv

Install dependencies

pip install catboost kaggle pandas numpy scikit-learn matplotlib seaborn

---

🛠️ Tech Stack

Tool	Purpose
pandas / numpy	Data wrangling and feature engineering
CatBoost	Primary gradient-boosted classifier
scikit-learn	Train/test split, OrdinalEncoder, ROC-AUC metric
matplotlib / seaborn	Visualisation
Kaggle API	Dataset download automation

---

📝 Notes & Limitations

• auto_class_weights='Balanced' is applied to handle the natural class imbalance in churn datasets (churned customers are typically a minority)
• drop_first=False was deliberately chosen for PaymentMethod one-hot encoding — all four payment categories carry distinct signal worth preserving despite the multicollinearity risk
• The InternetOutage and PhoneServiceOutage flags are the key novel features — they prevent the model from conflating service unavailability with active opt-out decisions
• LightGBM, XGBoost, and Random Forest benchmarks are planned as next steps to compare against the CatBoost baseline

---

👤 Author

Rony Soliman — Kaggle competition submission for Playground Series S6E3.