A data science project aimed at predicting a customer’s likelihood to subscribe to a bank term deposit based on demographic, campaign, and financial information. The project leverages machine learning models to identify critical factors influencing a customer's decision, allowing banks to optimize marketing strategies.
This project analyzes a bank marketing dataset to uncover insights into customer behavior and improve marketing campaign effectiveness. Machine learning models, including logistic regression, decision trees, and random forests, were applied to predict whether a customer is likely to subscribe to a term deposit.
- Source: A Portuguese banking institution’s marketing campaign data
- Features: Customer demographics, campaign interactions, subscription outcomes, and financial information.
├── data # Dataset files
├── notebooks # Jupyter notebooks for EDA and modeling
├── models # Trained model files
├── app.py # Streamlit app for interactive predictions
├── requirements.txt # Required libraries for the project
└── README.md # Project overview and instructions
- Handling Missing Values: Ensured data integrity by checking and handling missing values.
- Outlier Detection: Identified and managed outliers, especially in critical columns.
- Encoding: Converted categorical features to numerical values using one-hot encoding.
- Scaling: Standardized numerical features to improve model performance.
- Class Balancing: Used SMOTE to address imbalanced classes in the target variable.
EDA involved examining correlations and patterns in demographic data such as job, marital status, and education, which were found to impact subscription likelihood. Key findings include:
- Positive correlation between the number of contacts and the likelihood of subscription.
- Higher subscription rates among certain job categories and education levels.
Categorized features like age, duration, and balance to provide nuanced insights, enhancing the model’s predictive power by creating high-level representations of customer behavior.
Several models were trained and evaluated:
- Logistic Regression: Baseline model with high recall, suitable for identifying potential subscribers.
- Random Forest: Best model with the highest accuracy and F1 score, providing a balance of precision and recall.
- Decision Tree: An interpretable model, useful for understanding decision rules but less accurate than Random Forest.
| Model | Accuracy | Precision | Recall | F1 Score |
|---|---|---|---|---|
| Logistic Regression | 84.08% | 41.88% | 82.31% | 55.52% |
| PCA + Logistic Regression | 80.42% | 34.93% | 78.16% | 48.29% |
| Random Forest | 90.01% | 58.38% | 60.04% | 59.20% |
| Decision Tree | 87.12% | 45.71% | 53.88% | 49.46% |
An interactive Streamlit app link users with a simple interface for inputting customer details and generating predictions about the likelihood of subscription. The app also displays model accuracy and feature importances.