Anomaly Detection with Noisy Labels in Graphs

This research project focuses on improving how we detect fraudulent activities, such as suspicious transactions, in large financial transaction networks like Mastercard’s. Fraudulent transactions are rare and difficult to catch because they often look similar to regular transactions, and there are limited accurate data to rely on. To tackle this, we are using Graph Neural Networks (GNNs), which can analyze transaction networks in ways that traditional methods cannot by incorporating the graph structure in the model. This network view helps GNNs detect patterns of fraud that might otherwise go unnoticed. However, a big challenge is that some of the data we use to train these models is noisy, meaning it contains mistakes or inaccurate labels. These noisy labels can affect the model negatively, causing
it to learn incorrect patterns and reducing its ability to find real fraud.

Faculty Supervisor:

Pawel Pralat

Student:

Partner:

Mastercard

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

Toronto Metropolitan University

Program: