Fraud Detection in Financial Graphs

A significant amount of fraudulent activity tends to go unreported in reality, one of the major focuses for our team recently has been to develop robust GNNs that can perform anomaly detection on noisily labeled graphs. Noisily labeled anomalous data can reduce the model performance as it learns incorrect patterns during training, cause the model to overfit the noise in the labels as opposed to the underlying true anomalies, and can harm the message-passing mechanism of the GNNs.
Having a GNN based model that is robust to noisy labels can provide us with a significant uplift over the current production models. In research, there is limited work already done in this domain using GNNs. However, applying these methods to Mastercard’s data in a way that transfers well is a challenging task due to:
a) Heavy Class Imbalance b) Inductive – Model should perform well for out-of-distribution test set c) Scalability – the model should learn from over one billion noisily
labeled data d) Nature of Graph – Transaction graph is significantly different from other open source graphs used in existing research

Faculty Supervisor:

Euijin Choo

Student:

Partner:

Mastercard

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

University of Alberta

Program: