The characteristics of fraudulent financial transactions in the context of a graph model

Abstract

This thesis aims to answer the question if graph-based methods can be employed on available financial datasets with the purpose of detecting illicit financial activities. The data was gathered from three separate data sets – one being a synthetic PaySim dataset, the second one provided by Vesta in cooperation with the Institute of Electrical and Electronics Engineers (IEEE) and the third one related to Bitcoin transactions. In all cases, exploratory analysis is applied to attempt to gain an initial overview of the data sets and presumably to identify certain characteristics which can serve to find additional methods for fraud detection. The data are analyzed using graph-based approaches which allows for retrieving centrality metrics for different classes of nodes indicating if they are involved in fraudulent activity or not. The outcomes were examined using goodness of fit analysis and descriptive statistics measures to determine if there are differences between groups of observations. At a general level of metrics distribution in different observation classes, Mann-Whitney U test was employed. Finally, Louvain modularity was used to gather information regarding dense communities which can constitute fraud rings. The results of this study suggest that some of the methods presented in this paper can be useful, however, precise, non-anonymized data must be provided to prove their efficacy. In all our experiments, the centrality metrics did not perform well for predicting fraud. Without additional information on the entity making a transaction it is not possible to flag potentially suspicious nodes accurately.