DETECTING FRAUD BASED ON THE ANALYSES OF SOCIAL NETWORKS

Lovro Šubelj (2008) DETECTING FRAUD BASED ON THE ANALYSES OF SOCIAL NETWORKS. EngD thesis.

PDF Download (1214Kb)

Abstract

In this work we present a new system for detecting automobile insurance fraud. Here fraudsters stage car accidents and than issue fake claims so they can gain funds from insurance company. Specially interesting are groups of organised individuals made up of fraudulent drivers, chiropractics, garage mechanics, police officers and others. Our system focuses mainly on detecting such groups. In contrast to many other solutions, we use networks for the representation of data. Networks are possibly the most natural representation of accidents data, moreover, they provide straightforward formulation of complex relations between different entities. The last is crucial for detecting such frauds. Besides detecting key entities the system also provides visualisation of final results to the domain expert, as it is believed, that fully automatic automobile fraud detection is impossible. We use networks for that part of system as well, as the representation is clear and also relatively simple. Otherwise the system doesn't require labeled data set, it is relatively simple to implement, allows imputation of arbitrary domain knowledge and can be extended and improved in many ways. System was tested on real world data and the results were very good. We cannot estimate the efficiency in general, as the sample was not big and unrepresentative, but we are satisfied with the results. We also got a similar response from the analytics of slovenian insurance company. Key-words: fraud detection, social networks, graph theory, insurance.

Actions (login required)