Introduction
Graph Neural Networks, almost always abbreviated to GNNs, are a class of machine learning models designed specifically to analyse data that is structured as networks of connected entities. Where traditional models treat each record in isolation, GNNs incorporate the relationships between records, often surfacing patterns that would otherwise remain invisible.
In insurance fraud detection, GNNs offer particular value where organised fraud relies on connections between individuals, vehicles, addresses, suppliers, and claims.
What GNNs Mean (Plain English)
A graph, in machine learning terms, is a way of representing data as nodes (entities) and edges (relationships). A GNN is a neural network model that learns from the structure of the graph as well as the features of individual nodes.
In an insurance context, nodes might represent:
- Customers, drivers, and passengers
- Vehicles and their registrations
- Addresses, phone numbers, and email accounts
- Suppliers including repairers, medical providers, and lawyers
- Claims and policies
Edges represent the relationships between these — who lives at which address, which vehicles appear on which policies, which suppliers serve which claims, and so on.
Why GNNs Suit Insurance Fraud
Organised insurance fraud is fundamentally a network phenomenon. The same individuals, suppliers, or addresses appear across multiple claims, sometimes in slightly different forms to avoid detection. Traditional models that score each claim in isolation can miss these patterns even when the underlying signal is strong.
GNNs incorporate network context directly into the prediction. A claim that appears unremarkable on its own attributes may become highly suspicious in the context of its connections — the same telephone number appearing across several claims, or a supplier whose other claims have already been confirmed as fraudulent.
Typical Applications in Insurance
Common applications of GNNs in insurance fraud include:
- Identifying organised motor fraud networks across claims
- Surfacing supplier collusion patterns
- Detecting application fraud where multiple policies share underlying identifiers
- Identifying ring fraud across CTP, casualty, or property lines
- Prioritising investigation across high-volume claim portfolios
Key Requirements
GNNs depend on several foundational capabilities to work well:
- Strong entity resolution — without accurate matching, the network itself is unreliable
- High-quality, well-curated data covering customers, suppliers, claims, and identifiers
- Computing infrastructure capable of training and operating graph models at scale
- Specialist data science skills to design, train, and tune the models
- Explainability frameworks to support investigator and customer engagement
Common Challenges
Despite their potential, GNNs present real challenges:
- Explainability remains harder than for many traditional models, complicating engagement with investigators, customers, and regulators
- Performance can vary across types of fraud, depending on the strength of the network signal
- Model maintenance requires ongoing attention as fraud tactics evolve
- Integration with existing case management workflows requires careful design
Regulatory and Governance Considerations
Like all AI models used in insurance decision-making, GNNs must operate within Australia’s broader AI governance expectations. This includes APRA’s CPS 230 operational risk requirements, ASIC’s conduct expectations under the Corporations Act 2001, and the General Insurance Code of Practice 2020.¹,³,⁴
Explainability is particularly important. Where a GNN contributes to a decision affecting a customer, investigators must be able to understand and articulate the basis for any action taken. ASIC has noted that AI developments have outpaced the establishment of appropriate governance frameworks at many financial services firms.³
Role in a Layered Detection Strategy
GNNs are most valuable as part of a layered fraud detection strategy rather than as a replacement for other techniques. Combined with anomaly detection, supervised classification models, and traditional rule-based detection, GNNs add a network dimension that strengthens overall detection performance.
The combination — not any single technique in isolation — is what enables effective, explainable, and proportionate fraud detection at scale.
Related Topics
Network analysis in insurance fraud
Entity resolution and risk visibility
Anomaly detection in insurance fraud
Sources & further reading
¹ APRA — Prudential Standard CPS 230 Operational Risk Management
² APRA — Letter to Industry on Artificial Intelligence (30 April 2026)
³ ASIC — Report 798: Beware the gap — governance arrangements in the face of AI innovation
⁴ Insurance Council of Australia — General Insurance Code of Practice 2020
