📊 Telco Customer Churn Prediction - Machine Learning Pipeline

This project demonstrates a complete end-to-end machine learning workflow using the Telco Customer Churn Dataset. The objective is to build and compare multiple classification models to predict whether a customer is likely to churn, and identify the best-performing model using evaluation metrics and hyperparameter tuning.

---

📁 Dataset Used

Dataset: Telco Customer Churn

• Applied data cleaning and preprocessing
• Removed irrelevant features
• Performed encoding and scaling
• Conducted exploratory data analysis and visualization
• Trained and evaluated multiple ML models
• Applied hyperparameter tuning using GridSearchCV and RandomizedSearchCV

---

🎯 Objective

• Train and compare multiple machine learning models
• Evaluate models using Accuracy, Precision, Recall, and F1-Score
• Tune models using hyperparameter search techniques
• Analyze and select the best-performing model based on business objectives

---

🤖 Models Trained

• Logistic Regression
• Decision Tree
• Random Forest
• Support Vector Machine (SVM)
• K-Nearest Neighbors (KNN)
• Naive Bayes

---

📈 Visualizations Created

• Class distribution of the target variable
• Correlation heatmap
• Before vs After Scaling comparison
• Accuracy, Precision, Recall, F1-Score comparison (before and after tuning)

---

📊 Evaluation Metrics - Before Tuning

Model	Accuracy	Precision	Recall	F1-Score
Logistic Regression	0.8038	0.6485	0.5722	0.6080
Decision Tree	0.7050	0.4459	0.4519	0.4489
Random Forest	0.7797	0.6096	0.4759	0.5345
SVM	0.7910	0.6351	0.5027	0.5612
KNN	0.7527	0.5359	0.5187	0.5272
Naive Bayes	0.7356	0.5018	0.7513	0.6017

---

⚙️ Evaluation Metrics - After Hyperparameter Tuning

Model	Accuracy	Precision	Recall	F1-Score
Logistic Regression	0.8038	0.6485	0.5722	0.6080
Decision Tree	0.7754	0.5980	0.4733	0.5284
Random Forest	0.7910	0.6408	0.4866	0.5532
SVM	0.7989	0.6409	0.5535	0.5940
KNN	0.7719	0.5730	0.5561	0.5645
Naive Bayes	0.7356	0.5018	0.7513	0.6017

---

🏆 Best Performing Model

• Logistic Regression: Highest accuracy and balanced performance even after tuning. Interpretable and generalizes well.
• Naive Bayes: Highest recall — useful when the goal is to catch more churners (important for retention strategies).
• SVM & Random Forest: Achieved strong balance between precision and recall after tuning.

---

✅ Conclusion

This project successfully covered the entire machine learning pipeline:

• 📌 Preprocessing: Feature removal, encoding, and scaling
• 🧠 Modeling: Trained six classifiers and compared their performance
• ⚙️ Tuning: Improved results using GridSearchCV & RandomizedSearchCV
• 📉 Evaluation: Focused on meaningful metrics like recall and F1-score for imbalanced data
• 📊 Visualization: Enhanced interpretability and presentation of results

Key Takeaways:

• 📌 No single model fits all — selection depends on business goals (maximize recall vs precision)
• 📌 Naive Bayes is ideal when false negatives are costlier
• 📌 Logistic Regression is a solid, reliable baseline
• 📌 Hyperparameter tuning plays a vital role in boosting model performance

---

📂 Project Structure