Fraud Detection - Advanced SOTA Architecture

An advanced state-of-the-art fraud detection system that significantly outperforms traditional baselines by combining Graph Neural Networks (features), Deep Learning (TabNet), and GPU-accelerated Gradient Boosting into a robust meta-ensemble.

Status: Achieved 0.997+ PR-AUC, representing a 14%+ improvement over the original paper methodology.

---

Architecture

The system utilizes a multi-stage architecture that processes 6.3M+ transactions through specialized engines before making a final prediction.

graph TD
    A[Transaction Data] --> B[Feature Engine]
    A --> C[Graph Engine]
    A --> D[Filter Engine]
    
    B --> E[Feature Matrix]
    C --> E
    D --> E
    
    E --> F[LightGBM - GPU]
    E --> G[XGBoost - GPU]
    E --> H[TabNet - Deep Learning]
    E --> I[Isolation Forest - Unsupervised]
    
    F --> J[Meta-Learner]
    G --> J
    H --> J
    I --> J
    
    J[Logistic Regression] --> K[Final Fraud Prediction]

Loading

Key Components

1. Graph Engine (NetworkX)
   • Computes PageRank, Degree Centrality, and interaction ratios.
   • Identifies "fraud rings" and mule accounts based on network topology.
2. Deep Learning (TabNet)
   • Uses attention mechanisms to learn feature importance automatically.
   • Provides diversity against tree-based models.
3. GPU Acceleration (CUDA)
   • LightGBM and XGBoost trained on GPU for 7-8x speedup.
4. Robust Stacking Ensemble
   • Combines Supervised (LGBM, XGB, TabNet) and Unsupervised (Isolation Forest) models.
   • Meta-learner (Logistic Regression) optimizes the final probability.

---

Results & Performance

Rank	Model	PR-AUC	Precision	Recall	Improvement
1	SOTA Ensemble	0.997+	99%+	99%+	+14.1%
2	Stacking V1	0.9973	100%	69%	+14.1%
3	LightGBM (GPU)	0.9970	99%+	99%+	+14.0%
4	Improved RF	0.9961	100%	29.5%	+13.9%
5	Baseline RF	0.8743	6.3%	99.3%	-

---

Documentation & Reference

Detailed architecture and implementation details can be found in the docs/ directory:

• Advanced SOTA Architecture: Full breakdown of the final system including Graph features and TabNet.
• Stacking Ensemble Architecture: Details on the initial stacking approach and base model selection.

---

Quick Start

1. Requirements

• Python 3.9+
• CUDA-capable GPU (Recommended for TabNet/XGBoost)
• RAM: 16GB+ (32GB+ recommended for full graph features)

2. Setup

# 1. Setup environment
.\setup_environment.ps1

# 2. Download PaySim dataset from Kaggle
kaggle datasets download -d ntnu-testimon/paysim1 -p data/

# 3. Extract dataset
Expand-Archive data/paysim1.zip -DestinationPath data/

# 4. Start Jupyter
jupyter notebook notebooks/

3. Running the SOTA Model

Open notebooks/07_advanced_sota.ipynb to run the full pipeline including graph feature extraction and ensemble training.

---

Project Structure

fraud-detection/
├── configs/              # Model configurations (YAML)
├── docs/                 # Detailed architecture documentation
├── notebooks/            # Jupyter notebooks (01-07)
│   ├── 06_stacking_ensemble.ipynb  # Stacking implementation
│   └── 07_advanced_sota.ipynb      # Final SOTA implementation
├── src/                  # Source code modules
│   ├── models/           # TabNet, LGBM, XGBoost wrappers
│   ├── feature_engineering.py  # Graph & statistical features
│   └── sampling.py       # SMOTE-ENN handling
├── results/              # Figures and metrics
└── data/                 # Dataset (gitignored)

License

MIT License