| October 1, 2019
Project Overview
The goal of this project is to investigate some new techniques of machine learning applied to graphs. Several ways of modeling credit card transactions have been tested, together with algorithms.
The project explores whether representing the relationships between customers, merchants and transactions can provide useful information for detecting fraud in card payments, including in-store and online purchases. The students studied Node2Vec and GraphSAGE and compared simple, bipartite and time-sequenced graph representations, with weighted variants using transaction amounts.
Their experiments used Node2Vec to learn graph embeddings and then assessed these features in transaction classification, both alone and alongside the original transaction attributes. Adding the embeddings did not consistently improve the baseline classifier on the datasets tested. The work also explored combining the decisions of separate classifiers, examining the trade-off between reducing false positives and missing fraudulent transactions. Graph visualization and the interpretation of the learned representations formed another part of the investigation; further experiments with GraphSAGE were proposed as a continuation.