🔁 Customer Churn Prediction

Benchmarking of Machine Learning and Deep Learning models for customer churn prediction in the telecom industry, with automatic hyperparameter optimization via Optuna.

📌 About the Project
This project investigates the predictive power of five distinct approaches to identify customers at risk of churning (service cancellation), using a real-world telecom dataset. The full pipeline covers exploratory data analysis, preprocessing, feature engineering, model training, hyperparameter optimization, and comparative evaluation.

🧪 Models Evaluated
MLP
Random Forest
XGBoostGradient
TabTabular
TabPFN2

🗂️ Dataset

Source: Telco Customer Churn
Target: Churn — Yes / No

⚙️ Pipeline

1. DATA LOADING

2. PREPROCESSING
   ├── Convert TotalCharges to numeric (coerce errors)
   ├── Encode target variable: Churn → binary (0/1)
   ├── One-hot encoding of categorical features
   ├── Standardization with StandardScaler
   └── Train / Validation / Test split

3. EXPLORATORY DATA ANALYSIS └── Distribution plots for tenure, MonthlyCharges, TotalCharges

4. FEATURE ENGINEERING — Bucketing ├── MonthlyCharges → <22 | 22–68 | 68–105 | 105+
   ├── tenure → 0–3 | 4–69 | 70+
   └── TotalCharges → <200 | 200–2000 | 2000+

5. CLASS BALANCING
   └── RandomOverSampler

6. MODEL TRAINING + HYPERPARAMETER OPTIMIZATION (Optuna)

7. EVALUATION ├── Metrics: Accuracy, Balanced Accuracy, Precision, Recall, F1, AUC-ROC, KS Statistic
   ├── Confusion Matrix per model
   └── ROC Curve comparison

📊 Results