Circular Trade Fraud Detection with Graph Learning
Detecting circular trading rings in iron-dealer invoice data using cycle detection, Node2Vec embeddings and density-based clustering.
Problem
Circular trading — where invoices cycle through a ring of participants to inflate apparent turnover — is invisible to per-transaction rules, because every individual invoice looks ordinary. The signal only exists in the structure of the transaction graph.
Dataset
Iron-dealer transaction data, where each row is one invoice recording the seller, the buyer and the transaction value.
Method
- Build a multi-graph representation of the invoice data.
- Detect 2-cycles and 3-cycles and compute a suspicion score per node from cycle membership and trading amounts, using hash functions for efficient score storage across cycle nodes.
- Generate node embeddings with Node2Vec over undirected graphs derived from suspicion scores and cycle frequencies.
- Cluster with DBSCAN to isolate dense regions of potentially fraudulent nodes.
Results
The pipeline identifies dense clusters of nodes exhibiting anomalous trading behaviour consistent with circular trading patterns.
Limitations
Cycle detection is limited to 2- and 3-cycles, so longer rings are outside what this construction can surface. Outputs are unsupervised suspicion signals for investigation, not determinations of fraud.