Bank_Churn_Dataset

Bank Churn Prediction – End-to-End ML Project with Deployment

Deployed on Streamlit: View Live App

---

Table of Contents

1. Project Overview
2. Dataset
3. Data Exploration & Cleaning
4. Data Preprocessing
5. Model Selection
6. Hyperparameter Tuning
7. Final Model Performance
8. Predictions
9. Key Takeaways
10. Technologies Used

---

Project Overview

Predict whether a bank customer will churn (leave the bank) using historical customer data. The main goal is maximizing recall, since missing potential churners is costlier than giving false alarms.

---

Dataset

• Source: Kaggle Playground Series S4E1
• Files:
  • train.csv – 165,034 rows, 14 columns
  • test.csv – Test data for predictions
  • sample_submission.csv – Example submission format

---

Data Exploration & Cleaning

• Dropped irrelevant columns: id, CustomerId, Surname
• Categorical columns: Geography, Gender, HasCrCard, IsActiveMember
• Numerical columns: CreditScore, Age, Tenure, Balance, NumOfProducts, EstimatedSalary
• Removed outliers: Age > 85
• Converted categorical features:
  • Gender: Male = 1, Female = 0
  • HasCrCard & IsActiveMember → int

---

Data Preprocessing

• One-Hot Encoding for categorical variables
• StandardScaler for numerical features
• Train and test datasets aligned for consistent features

---

Model Selection

• Models evaluated:
  • Logistic Regression, Decision Tree, Random Forest, AdaBoost, Gradient Boost, XGBoost, KNN
• Metrics: Accuracy, Precision, Recall, F1-Score, ROC-AUC
• Top models: Gradient Boost and XGBoost
• Recall prioritized → XGBoost selected

---

Hyperparameter Tuning

• Gradient Boosting: n_estimators, learning_rate, max_depth, subsample
• XGBoost: n_estimators, learning_rate, max_depth, subsample, colsample_bytree
• Used GridSearchCV with StratifiedKFold (5 splits)
• Optimal XGBoost parameters:

XGBClassifier(
    colsample_bytree=1,
    learning_rate=0.05,
    max_depth=5,
    n_estimators=200,
    subsample=0.7
)

id,Exited
165034,0.036570
165035,0.144758
165036,0.083989

Final Model Performance (Test Split)

Metric	XGBoost
Accuracy	0.8663
Precision	0.739
Recall	0.568
F1-Score	0.642
ROC-AUC	0.8899

Key Takeaways

1. Business priority of maximizing recall ensures that potential churners are not missed
2. XGBoost outperforms other models for this classification task
3. Preprocessing (OHE + StandardScaler) ensures model consistency
4. Hyperparameter tuning improves model metrics (especially ROC-AUC and recall)
5. Deployment on Streamlit allows interactive, real-time predictions

Technologies & Libraries Used

Programming Language: Python 3.10

Libraries:

• pandas – data manipulation and analysis
• numpy – numerical computations
• scikit-learn – preprocessing, model selection, and evaluation
• xgboost – XGBoost classifier for modeling
• matplotlib & seaborn – data visualization
• streamlit - Web app deployment