This thesis describes the development of a fraud detection scheme for
car insurance customers, based only on information that is available at the
moment of underwriting. It explains how we manipulated raw anonymised
data and turned it into a graph, and how we used this graph to assign a
fraudulence score to each node. Finally, it evaluates the performance of this
score in identifying unknown fraudsters.
The results obtained in the thesis have been obtained by means of several
ad hoc optimised and parallel algorithms, which have been tested and run
on multiple HPC platforms.
