🔍 Customer Churn Prediction Using Machine Learning

This project aims to build a model to predict customer churn based on customer demographics, account information, and service usage using the Telco Customer Churn dataset from Kaggle.

📂 Dataset Link: Telco Customer Churn – Kaggle

---

📌 Objective

To build an accurate machine learning model that identifies customers likely to churn so that telecom companies can take proactive steps to retain them.

---

⚙️ Project Workflow

• Data cleaning and preprocessing
• Label encoding of categorical features
• Handled class imbalance using downsampling
  (SMOTE caused severe overfitting: Random Forest hit 99.8% train accuracy)
• Training with multiple classifiers
• Hyperparameter tuning using RandomizedSearchCV (also tried GridSearchCV)
• Model evaluation using test accuracy, F1-score, recall, and ROC AUC
• Deployment using Streamlit

---

🧠 Models Trained & Compared

Model	ROC AUC (Test)	Accuracy	Recall (Churn)	F1-score (Churn)
Random Forest ✅	0.8631	76%	86%	65%
XGBoost	0.8580	75%	82%	63%
CatBoost	0.8594	75%	83%	63%

🔎 Best model: Random Forest – highest ROC AUC, best balance of recall and F1-score

---

📈 Model Evaluation (Random Forest)

Train ROC AUC: 0.7921
Test ROC AUC: 0.8631
Accuracy: 76%
Recall (Churn): 86%
F1-score (Churn): 65%

---

🧪 Preprocessing Details

Step	Description
Categorical Encoding	Label Encoding via sklearn.preprocessing.LabelEncoder
SMOTE	Used SMOTE for upsampling, result gave better recall for non-churn but my aim was to better predict churn, so preferred downsampling
Class Imbalance	Used downsampling to balance churn/no-churn classes
Hyperparameter Tuning	Used RandomizedSearchCV with StratifiedKFold
Feature Engineering	Tried binning tenure into groups, but found negligible impact, so reverted

---

🚀 Streamlit Web App

A user-friendly web app is created using Streamlit to allow real-time predictions.

🔧 Features:

• Choose between Random Forest, XGBoost, or CatBoost
• Input form for customer attributes
• Shows prediction result and churn probability
• All models, encoders, and scalers are reused via pickle

📎 To Run the App:

# Step 1: Install dependencies
pip install -r requirements.txt

# Step 2: Run the app
streamlit run app.py

---

Link for streamlit app

Try mychurn predictor

---

🧰 Tech Stack

• Python 3.11
• Scikit-learn (Label Encoding, RandomForest, CV tools)
• CatBoost, XGBoost
• Streamlit – for web UI
• Pickle – for model serialization
• Pandas, NumPy, Seaborn, Matplotlib

---

🛠 File Structure

Churn_predictor/
├── app.py                          # Streamlit frontend
├── customer_churn_randomforest.pkl
├── customer_churn_catboost.pkl
├── customer_churn_xgboost.pkl
├── encoders.pkl
├── monthlycharges_scaler.pkl
├── totalcharges_scaler.pkl
├── requirements.txt
└── README.md

---

💡 Future Work

• Try to reduce overfitting
• Explore stacking ensemble (e.g. RF + XGB + CatBoost)
• Dockerize for scalable cloud deployment
• Add logging and analytics dashboard to track model predictions

---

🙌 Acknowledgements

• Telco Customer Churn Dataset – Kaggle
• Scikit-learn, Streamlit, XGBoost, CatBoost

---