Credit Card Fraud Detection System

Credit card fraud detection using Logistic Regression with EDA and class imbalance analysis

Overview

This project focuses on detecting fraudulent credit card transactions using machine learning.
I used Logistic Regression, a simple and interpretable classification algorithm, to identify whether a transaction is fraudulent or not.

A key challenge in this project is class imbalance, since fraud cases are very rare compared to normal transactions. To address this, I experimented with both unbalanced and balanced data.

---

Dataset

• Get the dataset I used in this system from: Source: Kaggle Credit Card Fraud Dataset
• The dataset is anonymized for privacy.
• Features V1–V28 are PCA-transformed and do not have direct real-world meaning.
• I used only Time and Amount as input features because they are interpretable.
• Target variable:
  • Class = 0 → Not Fraud
  • Class = 1 → Fraud

---

Notebooks in This Repository

This repository contains two notebooks:

1. Fraud Detection System (Main Notebook)

• Exploratory Data Analysis (EDA)
• Model training on unbalanced data
• Model evaluation
• Final deployed model
• Gradio-based user interface for predictions

2. Balanced Data Experiment

• Analysis of class imbalance
• Data balancing using SMOTE
• Training Logistic Regression on balanced data
• Performance comparison with the unbalanced model
• Shows how balancing affects accuracy and fraud detection
• SMOTE (Synthetic Minority Oversampling Technique) was used to generate synthetic fraud samples to balance the training data.

---

Exploratory Data Analysis

• The dataset is highly imbalanced, with fraud making up less than 1% of transactions.
• Visualizations used:
  • Bar chart for class distribution
  • Scatter plot (Time vs Amount)
  • Correlation heatmap
• These graphs help understand why fraud detection is a challenging problem.

The dataset contains 17,836 non-fraud transactions and only 81 fraud cases, making fraud less than 0.5% of all transactions.

---

Model Development

• Algorithm: Logistic Regression
• Train-test split: Stratified split to preserve class ratios
• Two approaches tested:
  • Training on unbalanced data
  • Training on balanced data using SMOTE

This allowed a direct comparison between both approaches.

---

Results

• The unbalanced model achieved very high overall accuracy but struggled with detecting fraud.
• The balanced model improved fraud detection but reduced overall accuracy.
• For this project, the unbalanced model was selected as the final system, while the balanced model is included as an experiment to demonstrate understanding of class imbalance.

---

Insights

• Unbalanced model achieves high overall accuracy but lower fraud recall.
• Balanced model improves fraud detection but reduces overall accuracy.
• Demonstrates trade-offs in class imbalance and model evaluation metrics.
• Highlights importance of precision, recall, and F1-score over raw accuracy in fraud detection.

---

Deployment

A simple Gradio-based GUI was built to simulate real-time fraud detection.
Users can input transaction details and instantly see whether the transaction is predicted as fraudulent or not. The interface accepts Time and Amount as inputs and outputs a fraud prediction.

You can try the interactive demo yourself, see DEMO.md in this repository for the link.

---

Technologies Used

• Python
• Pandas, NumPy
• Matplotlib, Seaborn
• Scikit-learn
• Imbalanced-learn (SMOTE)
• Gradio

---

Author

Amina, Information Technology university