Find the Fraud: Assessing Tax Non-Compliance in the IRS
summary
Applies graph neural networks to simulated IRS filing networks using a Python simulator, Neo4j-style graphs and PyTorch style GNNs, featurizing nodes with 16 form/country features plus 384 NLP embedding dimensions into 400-dimensional node vectors. Trains on 300 simulated graphs with typical graphs around 100–150 nodes, uses three convolutional layers with 64 hidden channels and dropout, and reports nearly 100% test accuracy on a binary risky vs clean classification without any real IRS data.
AI-written summary · about
chapters (9)
- 0:00Introduction and disclaimer
- 1:13Tax fraud in IRS networks
- 2:38Graph neural networks for fraud detection
- 6:13Graph convolutional neural networks
- 7:19Simulating realistic tax graphs
- 9:27Risky graphs and classification
- 11:15Natural language in node features
- 12:19Training the neural network
- 14:07Advantages of GNNs
related talks
12mCloud-Native Friendly Fraud Detection
26mAdvancing Financial Fraud Detection with Graph Dbs
45mApplications of Graph in Fraud Detection
52mModern Fraud Prevention Using Deep Learning
28mFraud Graph with Amazon Neptune and Machine Learning
28mA Quantum Leap In Fraud Detection Accuracy Using Graph-based Features
31mLet's do graph machine learning on Cloud
40mHow to Map Networked Devices Using Graph-Based AI to Preempt Cyber Threats