Transaction Fraud Detection Using GNNs and Tabular Models

This repository presents a comprehensive framework for fraud detection in financial transactions using both supervised and semi-supervised learning techniques. The dataset contains transaction details with labels indicating fraudulent or legitimate transactions.

Features

• Exploratory Data Analysis (EDA):

  • In-depth analysis of numerical and categorical features.
  • Temporal and categorical trend visualizations.
  • Identification of data imbalances and patterns.

• Supervised Learning:

  • Tabular models using engineered features.
  • Oversampling techniques (SMOTE) for handling class imbalance.
  • Comparison of models with and without feature engineering.

• Semi-Supervised Learning:

  • Graph-based models (Graph Attention Network (GAT) and GraphSAGE).
  • Hybrid loss combining classification and graph reconstruction objectives.
  • Alternative approaches: Autoencoders, Isolation Forest, Local Outlier Factor, and Gaussian Mixture Models.

• Visualization:

  • Temporal analysis of fraudulent transactions.
  • Network graph visualizations showing source-destination-agent relationships.
  • Distribution plots for transaction amounts and fraud probabilities.

Dataset

The dataset consists of financial transaction records with the following columns:

• date: Transaction timestamp in milliseconds.
• user: Unique user identifier.
• source_prefix, source_postfix: Attributes related to the transaction's source.
• dest_prefix, dest_postfix: Attributes related to the transaction's destination.
• agent: Categorical variable indicating the handler of the transaction.
• amount: Transaction value in the smallest monetary unit.
• status: Indicates the transaction's result (e.g., success or fail).
• label: Binary variable indicating fraud (1) or non-fraud (0).

Key Highlights:

• 106,036 rows, 10 columns.
• No missing values.
• Highly imbalanced, with ~3.5% fraudulent transactions.

Methods

1. Exploratory Data Analysis (EDA)

• Insights into data distribution, trends, and relationships.
• Fraud vs. non-fraud analysis for key features like amount, agent, and status.

2. Supervised Learning

• Models: Tabular Transformers, Neural Networks.
• Feature Engineering:
  • Transaction-specific features: is_high_risk_pair, fraud_rate.
  • Temporal features: is_night_high_risk.
• Oversampling: SMOTE for handling class imbalance.
• Evaluation Metrics: Precision, Recall, F1-Score.

3. Semi-Supervised Learning

• Graph-Based Models:
  • Graph Attention Networks (GAT).
  • GraphSAGE.
  • Hybrid loss combining node classification and reconstruction.
  • Note: Only 15% of the data is labeled for training.
• Other Methods:
  • Autoencoders.
  • Isolation Forest.
  • Local Outlier Factor.
  • Gaussian Mixture Models.

4. Visualization

• Temporal distribution of fraud cases.
• Network graphs of fraudulent transactions.
• Feature importance and correlation heatmaps.

Results

Method	Precision	Recall	F1-Score
Supervised (Engineered)	0.87	0.84	0.85
Supervised (SMOTE)	0.79	0.91	0.85
GAT (15% labeled data)	0.82	0.88	0.85
GraphSAGE (15% labeled data)	0.80	0.85	0.82

Note: Results may vary based on parameter tuning and dataset changes.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Acknowledgments

• Dataset Source: Provided during an interview task.
• Referenced Papers:
  • Semi-Supervised Learning on Graphs (Kipf et al.)