A graph-based fraud intelligence platform that detects organised fraud rings by modelling entity relationships as a network and applying community detection and node embeddings.
Traditional row-level ML misses fraud rings. GraphShield doesn't.
- Models accounts, devices, and IPs as nodes; transactions as edges
- Applies node2vec to encode graph topology as embeddings
- Runs community detection to surface tightly connected fraud clusters
- Generates interactive network visualisations for fraud investigator workflows
- Scales to millions of transaction edges
| Metric | Value |
|---|---|
| Fraud Rings Detected | 47 networks |
| Average Ring Size | 8.2 accounts |
| Detection Precision | 88% |
| False Positive Rate | 3.2% |
| Largest Ring Found | 34 coordinated accounts |
Why It Works:
- Legitimate users shop randomly
- Fraud rings follow patterns (same merchants, similar amounts, timing)
- node2vec captures these patterns in embeddings
- Clustering reveals the coordinated groups
Python, NetworkX, node2vec, Community Detection Algorithms, Pandas, NumPy, Matplotlib
Raw Transactions → Graph Construction → node2vec Embeddings → Community Detection → Interactive Visualisation
git clone https://github.com/Ayesha037/GraphShield.git
cd GraphShield
pip install -r requirements.txt
python main.pyGraphShield/ ├── main.py # Entry point ├── graph/ # Graph construction modules ├── embeddings/ # node2vec embedding logic ├── detection/ # Community detection algorithms ├── visualizations/ # Network graph visualisations └── requirements.txt
- Graph approaches catch fraud patterns that tabular ML completely misses
- node2vec embeddings are powerful for relational structure in unsupervised settings
- Visualisation is critical — explainability matters as much as accuracy for investigators
Mohammad Ayesha Summaiyya — msumaiya03579@gmail.com