Financial Fraud Detection Using Graph Data Science

The problem

PaySim simulates mobile money transactions with built-in rule-based fraud detection. I modeled 6.3M transactions across 9M+ accounts as a 15.4M-node Neo4j graph to find patterns invisible at the individual-transaction level.

The existing rule-based system caught only 16 of 8,213 fraud transactions — a 99.81% miss rate concealing over $12B in fraudulent transfers. That is not a tuning problem. It is a representation problem.

Approach

I modeled clients, merchants, and money flows as a bipartite transaction graph in Neo4j, then applied Graph Data Science algorithms independently:

• Louvain community detection to surface dense clusters of suspicious accounts
• PageRank to identify peripheral and hub accounts in the transfer network
• In-Degree Centrality to flag sink accounts that accumulate fraudulent transfers
• Weakly Connected Components to isolate disconnected fraud subgraphs

Findings were cross-validated with Cypher queries across 9M+ client accounts so no single algorithm's bias would drive the conclusion.

Key finding

Fraud hides in network structure, not individual row flags. Louvain identified 2,660 fraud account pairs unsupervised (~65% recall), and three independent methods converged on the same sink accounts rules never saw.

Results

• Only 16 of 8,213 fraud transactions caught by the rule-based system (99.81% miss rate)
• 2,660 fraud account pairs recovered via Louvain community detection (~65% unsupervised recall)
• Independent confirmation of sink accounts through Cypher, PageRank, and In-Degree Centrality
• $12B+ in fraudulent transfers made visible through graph structure alone

Implications

Rule-based systems remain useful for known fraud patterns, but they systematically miss relationship-driven schemes. For financial crime detection, graph data science isn't a nice-to-have — it is how you find what tabular flags cannot see.

Interested in similar work? Get in touch.