Anomaly Detection in Credit Card Fraud Transactions

---

Table of Contents

• Project Overview
• Dataset
• Installation
• Project Steps
  • 1. Data Preprocessing
  • 2. Feature Engineering
  • 3. Data Visualization
  • 4. Model Building
  • 5. Threshold Tuning
  • 6. Evaluation
• Results
• Conclusion

---

Project Overview

Credit card fraud detection is a critical task in financial institutions. The goal of this project is to build a system capable of identifying fraudulent transactions in a dataset where such cases are rare (0.17%). We use a multivariate normal distribution to model the probability density of legitimate transactions and classify those with lower densities as fraudulent.

---

Dataset

The dataset used in this project contains credit card transactions made by European cardholders in September 2013. It is highly imbalanced, with only 0.172% of transactions being fraudulent.

• Download the dataset from Kaggle

---

Installation

1. Clone this repository:

   git clone https://github.com/your-username/credit-card-fraud-detection.git
   cd credit-card-fraud-detection

2. Install the required libraries:

   pip install numpy pandas matplotlib seaborn scikit-learn tqdm psutil

3. Download the dataset from Kaggle and place it in the project directory.

---

Project Steps

1. Data Preprocessing

• Dataset Loading: Load the dataset and split it into two main categories: legitimate transactions (Class = 0) and fraudulent transactions (Class = 1).

• Splitting Data: The legitimate transactions are further divided into training, validation, and testing sets. The fraudulent data is split into validation and testing sets only.

• Merging Validation and Testing Sets: Combine the legitimate and fraudulent data for validation and testing.

---

2. Feature Engineering

• Time Feature Transformation: Decompose the Time feature into Day, Hour, Minute, and Second to extract more meaningful patterns.

• Amount Transformation: Apply a log transformation to the Amount feature to reduce skewness and stabilize variance.

---

3. Data Visualization

• Histogram and KDE Plots: Visualize the distribution of key features like Time, Hour, and Amount_transformed to understand the underlying data patterns.

Time - Hour

Amount - Transformed_amount

Features-Selection

---

4. Model Building

• Multivariate Normal Distribution: Fit a multivariate normal distribution to the training data by calculating the mean and standard deviation of each selected feature.

• Probability Density Calculation: Calculate the joint probability density function for the features, and classify transactions as fraudulent if their density falls below a certain threshold.

---

5. Threshold Tuning

• Tuning the Threshold: Iterate over a range of alpha values and select the one that optimizes the F2-score, focusing on reducing false negatives due to the imbalanced dataset.

---

6. Evaluation

• Confusion Matrix: Generate the confusion matrix to visualize the model's performance in classifying transactions.

• Performance Metrics: Calculate accuracy, precision, recall, F1-score, F2-score, and MCC to evaluate the model.

---

Results

The optimal threshold value was approximately 3.87 x 10^-19, resulting in an F2-score of 0.836 on the validation set and 0.815 on the test set. The confusion matrix illustrates the classification performance.

---

Conclusion

This anomaly detection system effectively identifies fraudulent transactions by modeling legitimate transactions using a multivariate normal distribution. Key takeaways include:

• Feature Selection: The choice of meaningful features significantly impacts model performance.
• Threshold Optimization: Proper tuning of the decision threshold is essential for handling imbalanced datasets.
• Real-World Application: The model achieves a high F2-score, making it suitable for deployment in real-world fraud detection systems.

---

Credits

This project uses the credit card fraud dataset from Kaggle.

Feel free to contribute to this project by submitting pull requests or suggesting improvements.

---