Why graph algorithms catch fraud that rule-based systems miss

Fraud lives in the graph, not the row

When I loaded 6.3 million PaySim transactions across 9M+ accounts into a 15.4M-node Neo4j graph, the first thing that struck me wasn't the volume — it was how invisible fraud became when you only looked at tabular flags.

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

Rule-based systems ask "does this row look bad?" Graph algorithms ask "who is this account connected to, and does that neighborhood behave like fraud?"

What the graph made visible

I modeled the money flows as a bipartite network and ran several Graph Data Science algorithms independently:

• Louvain surfaced dense communities of accounts transacting in suspicious patterns
• PageRank highlighted peripheral and hub roles in the transfer structure
• In-Degree Centrality flagged sink accounts that accumulate fraudulent transfers
• Weakly Connected Components isolated disconnected fraud subgraphs

The critical validation step: these methods independently pointed at the same structural signatures across 9M+ client accounts. When Cypher queries, PageRank, and centrality converge, you have a finding — not an artifact of one algorithm.

What Louvain revealed

Louvain identified 2,660 fraud account pairs unsupervised — about 65% recall. No single-transaction rule would ever flag those clusters, because the fraud signal lived in the relationships between accounts, not in any one row's feature values.

The takeaway

When relationships matter — and in financial crime they almost always do — graph data science isn't optional. It's the difference between catching 0.19% and catching 65%. Rules still have a place for known patterns. Graphs find the patterns you don't yet know how to write rules for.

Want to discuss this work? Get in touch.