Source-linked AI summary
An expert system for detecting automobile insurance fraud using social network analysis
Lovro Šubelj, Štefan Furlan, Marko Bajec
TL;DR
The paper addresses detection and investigation of collaborating automobile insurance fraudsters under limited and unlabeled data. It proposes a network-based expert system using the IAA algorithm to assess relational entities, and evaluates the prototype on real-world data. The results indicate that fraud can be efficiently detected and that appropriate data representation is vital.
Problem
Collaborating automobile insurance fraud is difficult to detect because groups, relational information, and appropriate data representation must be handled under practical data limitations.
Method
The system represents entities and their relations as networks and uses the Iterative Assessment Algorithm to assess them without an initial labeled data set.
Results
The prototype correctly classified almost 90% of fraudsters and over 85% of nonfraudsters, with only 5 fraudsters missed.
Takeaways & Limitations
Automobile insurance fraud can be efficiently detected with the proposed system, and appropriate data representation is vital for detecting related entities.
Takeaways & Limitations
The system cannot autonomously make final fraud assessments; domain experts or investigators must investigate results and determine further actions.
Abstract
from arXiv · showhide
The article proposes an expert system for detection, and subsequent investigation, of groups of collaborating automobile insurance fraudsters. The system is described and examined in great detail, several technical difficulties in detecting fraud are also considered, for it to be applicable in practice. Opposed to many other approaches, the system uses networks for representation of data. Networks are the most natural representation of such a relational domain, allowing formulation and analysis of complex relations between entities. Fraudulent entities are found by employing a novel assessment algorithm, \textit{Iterative Assessment Algorithm} (\textit{IAA}), also presented in the article. Besides intrinsic attributes of entities, the algorithm explores also the relations between entities. The prototype was evaluated and rigorously analyzed on real world data. Results show that automobile insurance fraud can be efficiently detected with the proposed system and that appropriate data representation is vital.
1. Introduction
Automobile insurance fraud includes collaborative groups that are difficult to detect with manual investigation and unsuitable data representations. The proposed network-based expert system uses the IAA algorithm to assess related entities without labeled data and supports subsequent investigation.
- 1. Introduction: Collaborating automobile insurance fraudsters create groups that are difficult to detect, while organized groups represent the majority of revenue leakage.
- 1. Introduction: Approximately 20% of insurance claims are estimated to be fraudulent, but many go unnoticed because investigations are usually manual and data representation is often inappropriate.
- 1. Introduction: Fraud detection must handle few fraudulent cases, scarce labeled and unlabeled data, emerging fraud types, and the need for expert-led investigation.
- 1. Introduction: The proposed system detects groups of collaborating fraudsters and their connecting accidents rather than isolated fraudulent entities.
- 1. Introduction: Networks represent relations among entities, enabling complex analysis, group detection, and visualization for subsequent investigation.
- 1. Introduction: The IAA-based expert system requires no initial labeled data set and allows domain knowledge to be incorporated for practical detection of new fraud types.
2. Related work
Prior fraud-detection research uses diverse statistical, machine-learning, and network methods, but many approaches depend on labeled data or target structures rather than entities. The paper positions IAA as a network-based ranking method designed for unlabeled, smaller fraud-detection data sets.
- 2. Related work: Existing fraud-detection research includes data mining, machine learning, neural networks, fuzzy logic, genetic algorithms, support vector machines, regression, trees, and statistical methods.
- 2. Related work: Network-analysis approaches detect suspicious structures, centrality patterns, or random-walk anomalies, but some target networks rather than suspicious entities or suit larger networks.
- 2. Related work: Relational classification and propagation methods commonly rely on initial or partial labeled data, limiting their suitability for insurance-fraud detection.
- 2. Related work: IAA assesses and ranks entities without an initial labeled data set, differing from inference-oriented classification methods and simplifying related propagation approaches.
- 2. Related work: The proposed approach uses networks, does not demand labeled data, and is intended for smaller data sets and complex relational analysis.
3. (Social) networks
The paper introduces graphs and networks as representations of vertices, edges, labels, directions, multiplicity, and connectivity. These concepts support analysis of relational structure, including distances, communities, and edge betweenness.
- 3. (Social) networks: A graph consists of vertices and edges, with networks providing a natural representation for domains involving relations or patterns among entities.
- 3. (Social) networks: Directed graphs use ordered vertex pairs for edges, while multigraphs allow multiple edges and loops; labeled graphs attach information to vertices and edges.
- 3. (Social) networks: The collision network depicts participants and vehicles as different vertex types, with directed edges representing collisions between vehicles.
- 3. (Social) networks: Vertex degree counts incident edges, while distance counts edges along a route and graph diameter is the maximum distance between vertices.
- 3. (Social) networks: Graphs may be cyclic or acyclic, and a connected graph can be partitioned into components; the discussion assumes a single connected component for later definitions.
- 3. (Social) networks: Communities are densely connected vertex subsets with relatively few edges between groups, and edge betweenness counts geodesics passing along an edge.
4. Expert system for detecting automobile insurance fraud
The expert system represents automobile-insurance fraud as networks, filters suspicious components, and then scores entities within them using the Iterative Assessment Algorithm. Its staged analysis combines structural indicators, relational information, and visualization while retaining domain-expert investigation for final assessment.
- The four-module framework constructs networks, detects suspicious components, identifies suspicious entities, and visualizes results for subsequent investigation.The first three modules perform construction, component screening, and entity assessment; the fourth supports expert review.
- 4.2. Suspicious components detection: Suspicious components are screened using structural properties before entity-level assessment, reducing the search space for later analysis.The component stage focuses on properties such as diameter and cycles, while the next stage locates key entities within retained components.
- 4.1. Representation with networks: Networks represent relational attributes and enable detection of collaborating fraudsters that flat, entity-isolated data cannot adequately capture.The system uses network structure to model relations between entities and to support group detection and assessment visualization.
- 4.2. Suspicious components detection: A majority-style component screen is limited because its fixed threshold ignores indicator prevalence and treats all indicators as equally important, while labeled data are scarce.The article identifies three weaknesses: the n/2 threshold may be inappropriate, component frequencies are ignored, and indicator weights are equal.
- 4.3. Suspicious entities detection: IAA assigns participant suspicion scores by combining intrinsic attributes with related entities’ assessments and iteratively propagating information through the network.The assessment model uses neighboring scores, intrinsic attributes, and relational attributes; iterating to a fixed point can overfit, whereas running for the component diameter performs better empirically.
- 4.4. Final remarks: The prototype’s empirical evaluation reports satisfactory performance, while the system’s outputs still require investigation by a domain expert or investigator.The paper explicitly rejects fully autonomous fraud detection in practice and presents visualization as support for subsequent analysis.
5. Evaluation with the prototype system
The prototype was evaluated on merged labeled and unlabeled collision data using PRIDIT for suspicious-component detection and IAA for participant assessment. Results examined model variants, iteration settings, anomaly-detection baselines, and the importance of preserving network structure.
- 5.1. Data: The evaluation used 3,451 participants involved in 1,561 Slovenian collisions from 1999–2008, combining labeled and unlabeled data.The labeled set focused on previously investigated fraudsters and removed almost all isolated collisions; the larger set supported PRIDIT analysis.
- 5.2. Results: The system used PRIDIT to identify suspicious components and IAAmean to assess fraudulent participants, reporting classification metrics under a cost-minimizing threshold.Reported metrics included accuracy, recall, precision, specificity, F1, and AUC; AUC did not require a suspicion threshold.
- 5.2. Results: 0.8872, 0.9145, 0.8942, and 0.9228 are the reported comparison values for the analyzed assessment models.The values are presented with the model-comparison analysis, whose table reports different IAA assessment models after PRIDIT analysis.
- 5.2. Results: Dynamic iteration counts were used because IAA performance depends on iteration number, and the analysis reports superior performance with dynamically selected iterations.The dynamic setting uses each component’s diameter together with the average diameter across components.
- 5.2. Results: The analysis found that incorporating domain-based factors improves the flexibility of IAA for detecting newly recognized fraud types.The paper attributes this flexibility to combining domain knowledge with improved factor assignment.
- 5.2. Results: Flat-data clustering produced no results comparable to the network-based results, supporting the importance of relational data representation.The comparison examined standard unsupervised methods including k-means and hierarchical clustering.
6. Discussion
The discussion finds that network-based IAA can detect substantial automobile insurance fraud, with performance shaped by domain knowledge, iteration settings, and component detection. It also identifies practical benefits and boundaries for deployment and broader application.
- Almost 90% of fraudsters and over 85% of nonfraudsters were correctly classified, with only 5 fraudsters missed by the prototype.The resulting recall was particularly high, which the authors identify as essential for fraud detection.
- IAAraw could already be sufficient in some circumstances, domain-expert factors significantly improved it, and IAAmean further improved both models.The factor analysis confirmed the importance of expert knowledge for final assessment.
- Early IAA iterations improved performance, but excessive iterations over-fit the model and reduced final performance, especially for AMmean.The authors recommend evaluating models using this performance decrease and note that the optimal iteration count depends on the data set and model.
- PRIDIT-based component detection combined with IAA outperformed centrality measures, while network representation enabled simple approaches to detect a reasonably large portion of fraud.Without component detection, IAA and centrality measures showed no significant difference; with PRIDIT, IAA dominated the alternatives.
- Hierarchical component detection can simplify data collection because relational attributes can identify components before complete intrinsic data are available.This is important in sensitive domains where complete participant data often cannot be obtained.
- The framework and IAA algorithm are presented as applicable beyond automobile insurance to other relational domains involving groups of related entities.The proposed framework remains structured hierarchically to support practical deployment.
7. Conclusion
The proposed network-based expert system detects and investigates groups of automobile insurance fraudsters without labeled data, while supporting adaptation across fraud types and relational domains. Evaluation indicates efficient detection and emphasizes that proper data representation is vital; future work targets unsupervised assessment and broader applications.
- The system efficiently detects automobile insurance fraud groups, with evaluation showing that proper data representation is vital.
- It avoids labeled datasets by incorporating domain-expert knowledge, enabling adaptation to newly recognized fraud types.
- The framework can help investigators detect and investigate fraud faster and more efficiently, while remaining easy to implement.
- The approach is also applicable to fraud detection in other relational domains.
- Future work will examine alternative, including nonlinear, assessment models for the IAA algorithm.
- Future work will make IAA unsupervised and investigate applications in other domains.