You can deploy this model and start earning money today!

erkhan7/
quantum

Views

No views yet

FINQOR Fraud Detection ML Pipeline

This repository contains the completely autonomous machine learning pipeline for the FINQOR platform, designed to detect fraudulent financial schemes using graph analysis and gradient boosting.

Requirements & Installation

A requirements.txt file has been provided to easily install the required libraries.

Open your terminal, navigate to this directory, and run:

pip3 install -r requirements.txt

This will install the correct versions of robust, open-source AI and data science libraries:

pandas & numpy (for tabular data manipulation)
networkx (for graph construction and feature extraction)
scikit-learn (for the LightGBM/GBDT tree-based ML model)
matplotlib (for network visualization)
joblib (for saving/loading the trained model weights)

How to Run the Model and Pipeline

To train the model through the training data and run inference on the transactions, simply execute the main script:

python3 finqor_pipeline.py

What happens when you run this script?

The pipeline executes the following 5 stages autonomously:

Synthetic Dataset Generation: The script builds training data simulating transactions between natural persons (Person_XX), Limited Liability Partnerships (TOO Company_XX), Joint-Stock Companies (JSC Company_XX), and Individual Entrepreneurs (IP Company_XX) in Kazakhstan (Almaty, Astana, Shymkent, etc.). It injects real fraud topologies (Fan-out, Transit chains, Cyclic loops). This generates nodes.csv, edges.csv, and labels.csv.
Graph Construction & Feature Extraction: The script builds a directional NetworkX graph. It calculates powerful structural features for every entity (Fan-in ratio, Fan-out ratio, PageRank, averages of incoming/outgoing funds).
ML Training & Inference: Using the HistGradientBoostingClassifier (a native Scikit-Learn implementation of LightGBM that is robust natively on Mac/Windows without libomp errors), the model is trained on these graph features. After training, the weights are saved to finqor_model.pkl. The model then scores every node with a Fraud Probability (Risk Score between 0.0 and 1.0).
Report Generation: The pipeline filters the top offenders and writes report.txt, documenting exactly why these nodes were identified (e.g., "Fan-out behavior detected").
Visualization: The script generates an image of the fraudulent network clusters (finqor_graph.png), with nodes colored deeply red for high risk.

Production Explainability (The "Why")

When presenting FINQOR, note that this Python script runs 100% locally and securely. We only pass the aggregated Risk Scores and structural flags (Fan-out > 0.8) from this script to our LLM (GPT) to generate human-readable explanations. Raw transaction IPs, names, and logs never touch external AI servers.

erkhan7/quantum