The Impact of Role Mining on Link Prediction in Financial Graph Networks

Abstract

This thesis investigates the effects of role mining on learning-based link prediction models in financial graph networks. Link prediction is utilised in the financial domain for predicting transactions and finding anomalies. This benefits the effective recognition of fraud and the detection of potential financial opportunities. Role mining has yielded positive results in other domains for improving link prediction models. However, this research finds that role mining does not significantly improve link prediction in financial graph networks. Role mining likewise does not seem complementary to other graph-based features, such as node centrality features and node similarity features. The features were provided in various combinations to a random forest algorithm. Whereupon the performance was evaluated based on the recall score over various thresholds. The findings in this paper contribute to a better understanding of role mining in financial networks and its effects on link prediction. It opens the way for further research into the effective use of appropriate feature combinations for link prediction in financial networks, in which role mining could potentially be instrumental.