Credit-Card-Fraud-Detection

Table of Contents

1. Introduction
2. Data Exploration and Analysis
   • Loading the Data
   • Exploratory Data Analysis (EDA)
   • Feature Engineering
3. Modeling
   • Model Selection
   • Resampling Techniques
   • Threshold Selection
   • Cross-Validation
4. Implementation
   • File Structure
   • credit_fraud_train.py
   • credit_fraud_test.py
   • credit_fraud_utils_data.py
   • credit_fraud_utils_eval.py
5. Reports
6. model.pkl File
7. Conclusion

Introduction

This project focuses on the binary classification problem of detecting credit card fraud. The goal is to build a robust model that can accurately classify transactions as fraudulent or legitimate.

Data Exploration and Analysis

Loading the Data

• The dataset is loaded using Pandas.
• Initial inspection includes checking the dataset shape, column types, and summary statistics.

Exploratory Data Analysis (EDA)

• Missing Values: Analyzed and handled any missing values in the dataset.
• Data Distribution: Visualized the distribution of key features and the target variable.
• Class Imbalance: Checked for imbalance in the target classes (fraud vs. non-fraud).
• Correlation Analysis: Investigated correlations between features.

Feature Engineering

• Scaling/Normalization: Applied scaling techniques to ensure features are on a similar scale.
• Class Imbalance Handling: Techniques such as SMOTE, undersampling, or oversampling were applied to address class imbalance.

Modeling

Model Selection

• Algorithms Used: Evaluated models including Logistic Regression, Random Forest, Gradient Boosting, AdaBoost, Neural Networks, and Voting Classifiers.
• Evaluation Metrics: Metrics such as accuracy, precision, recall, F1-score, PR-AUC, and ROC-AUC were used to assess model performance.

Resampling Techniques

• Techniques: Resampling techniques such as SMOTE, undersampling, and oversampling were applied to address class imbalance.
• Reports: Detailed reports for each resampling technique (SMOTE, undersampling, oversampling) are saved in the Report folder.

Threshold Selection

• Best Threshold: The best classification threshold was selected based on F1-score and other evaluation metrics to balance precision and recall.

Cross-Validation

• Cross-Validation: Cross-validation was used to validate the models' performance and to avoid overfitting.

Implementation

File Structure

The project consists of the following files:

• credit_fraud_train.py: Main script for training models based on user input via argparse.
• credit_fraud_test.py: Script for testing the trained model on a test dataset.
• credit_fraud_utils_data.py: Utility functions for data loading and preprocessing.
• credit_fraud_utils_eval.py: Utility functions for model evaluation and threshold selection.

credit_fraud_train.py

• Purpose: Script for training multiple models and selecting the best one based on evaluation metrics.
• Features:
  • Loads and preprocesses training and validation data.
  • Applies resampling techniques.
  • Trains models.
  • Evaluates models and saves the best-performing model along with the optimal threshold.

credit_fraud_test.py

• Purpose: Script for testing the saved model on a test dataset.
• Features:
  • Loads and preprocesses test data.
  • Loads the trained model and applies it to the test data.
  • Generates evaluation reports including classification metrics and ROC-AUC score.

credit_fraud_utils_data.py

• Purpose: Contains functions for data loading, cleaning, and preprocessing.
• Key Functions:
  • load_data(): Loads the dataset.
  • preprocess_data(): Preprocesses the data (e.g., handling missing values, scaling).

credit_fraud_utils_eval.py

• Purpose: Contains functions for evaluating models and selecting the best threshold.
• Key Functions:
  • evaluate_model(): Evaluates the model using various metrics.
  • find_best_threshold(): Finds the optimal threshold for classification.
  • generate_report(): Generates detailed reports for each model and resampling technique.

Reports

• Location: Detailed reports for each resampling technique (SMOTE, undersampling, oversampling) are stored in the Report folder.
• Content: Each report includes metrics such as F1-score, PR-AUC, and the best threshold for different models.

models

• The model.pkl file contains a dictionary with:
  • The trained model.
  • The best classification threshold.
  • Any other necessary information for model evaluation.

Conclusion

• Summary: The project successfully identifies the best model for detecting credit card fraud, balancing precision and recall across various resampling techniques.
• Future Work: Possible improvements include exploring additional features, advanced ensemble methods, and real-time fraud detection.