Syndicate Detection using Graph Neural Networks (GNN)

This project implements a sophisticated Fraud Detection System (FDS) enhanced with Graph Neural Network (GNN) techniques, tailored to identify fraudulent activities within digital ecosystems, specifically within online gambling and digital financial transactions in Indonesia.

Overview

With the rise of complex and organized fraud syndicates in the digital ecosystem, traditional rule-based fraud detection methods fall short. This project introduces a GNN-based approach to model relationships and behaviors within the transaction network, offering a powerful tool to detect fraud syndicates by analyzing patterns, connections, and centrality measures.

Features

• Graph Scaling: Handles large-scale datasets, with the capability to process bulk inputs of up to 10,000 users.
• Node Embeddings: Uses GNN to generate embeddings for nodes (users, transactions), which can then be clustered and analyzed.
• Fraud Detection: Utilizes a combination of graph analysis and machine learning to distinguish between fraudulent and non-fraudulent users.
• Interactive Visualization: Provides a network graph visualization with search functionality to zoom in on specific user IDs, displaying key metrics such as risk score, syndicate score, and transaction details.
• Real-time Analysis: Supports real-time detection with the ability to search and analyze user-specific data on demand.

Installation

1. Clone the repository:

   git clone https://github.com/natgluons/Syndicate-Indication-using-Network-Graph-Analytics.git

2. Navigate to the project directory:

   cd Syndicate-Indication-using-Network-Graph-Analytics

3. Install the necessary Python packages:

   pip install -r requirements.txt

Usage

1. Train the Model

Before using the system for detection, you need to train the model using the provided train_hgt_model.py script:

python train_hgt_model.py --data_dir ./realdata --model_dir ./model

2. Perform Fraud Detection

Once the model is trained, use the hgt_kmeans_riskscore_graph.py script to perform fraud detection on a new set of data:

python hgt_kmeans_riskscore_graph.py --data_dir ./realdata --model_path ./model/trained_hgt_model_realdata.pth --output_dir ./result

3. Visualize the Network Graph

Use the index.html and network_graph.js files to visualize the network graph. This provides an interactive interface where you can search for specific user IDs and zoom in to view their relationships and metrics:

1. Start a local server (for example, using Python's http.server):

   python -m http.server

2. Open index.html in your web browser.

4. Search and Analyze Specific Users

• Enter the user ID in the search bar and click "Search" to zoom in on the node representing that user.
• The node will be highlighted, and a tooltip will display detailed information about the user's inbound/outbound transactions, risk score, and syndicate score.

Data Description

• Nodes: Represent users or entities (e.g., merchants, customers).
• Edges: Represent transactions between entities.
• Features: Various features are extracted from nodes and edges, such as transaction amounts, account types, and other demographic attributes.

Methodology

• Heterogeneous Graph Transformer (HGT): Used to handle the complex relationships within the heterogeneous graph of transactions.
• K-means Clustering: Applied to cluster similar behaviors together, separating fraudulent and non-fraudulent entities.
• Syndicate Score: A composite score calculated using in-degree, out-degree, and betweenness centrality measures to identify potential syndicate activities.