Credit Card Behavior Score Analysis

Project Overview

This project aims to predict credit card behavior scores, such as the likelihood of default, using machine learning. Here, I utilized a dataset of credit card users and applied a series of data science techniques to build a robust predictive model. These techniques include data cleaning, feature engineering, feature selection, handling class imbalance, model training, and evaluation.

Key Features

• Data Preprocessing: Cleans the data by handling missing values, converting data types, and engineering new features from existing ones.
• Feature Selection: Selects important features using L1-based feature selection with LightGBM and correlation analysis.
• Class Imbalance Handling: Addresses class imbalance in the target variable (bad_flag) using the SMOTE technique.
• Model Training: Trains various machine learning models, including LightGBM, Random Forest, Gradient Boosting, and Logistic Regression, and evaluates their performance.
• Hyperparameter Tuning: Optimizes the LightGBM model's hyperparameters using RandomizedSearchCV for improved accuracy.
• Ensemble Modeling: Combines predictions from multiple models to potentially enhance prediction accuracy.
• Model Evaluation: Employs metrics like AUC-ROC, classification report, and confusion matrix to evaluate model performance.
• Model Persistence: Saves the trained model as a pickle file (credit_risk_model.pkl) for future use.

Dataset

The project utilizes two datasets:

• Dev_data_to_be_shared_mini.csv: The training dataset.
• validation_data_to_be_shared_mini.csv: The validation dataset.

Files

• Dev_data_to_be_shared_mini.csv: Training dataset.
• validation_data_to_be_shared_mini.csv: Validation dataset.
• predictions.csv: Predictions on validation data from the best model.
• credit_risk_model.pkl: Saved model file.

Results

The project achieved an AUC-ROC score of [0.7548] on the validation set using the [LightGBM] model.

Instructions

To run this analysis:

1. Clone this repository to the coding environment.
2. Upload the required datasets (Dev_data_to_be_shared_mini.csv, validation_data_to_be_shared_mini.csv) to your Colab environment.
3. Run all cells in the notebook.ipynb notebook.

Further Development

• Further refine the feature engineering and selection process.
• Explore other model types and ensemble methods.
• Implement more advanced error analysis techniques.
• Deploy the model as a web service for real-time predictions.

Author

[Sharvil Acharya]

Acknowledgements

• [Scikit-Learn]