Exploring Graph Neural Networks for Fraud Detection: A Deep Dive into Modeling Considerations

I am excited to share my reflections on a topic that has been on my mind during the LGFDS side project. Your feedback on these thoughts would be highly appreciated. Feel free to share your perspectives!

Topic: Modeling Considerations for GNN Message Passing Algorithm Targets

Subtopic 1: Dividing into Two Graph Types

One key consideration in our project is whether it is reasonable to split our modeling into two distinct graph types, namely graph deep learning and graph analytics.

I have expressed concerns about mixing these graph types, as it might introduce noise into the performance of graph deep learning. I would love to hear your thoughts on this matter.

I have attached two figures to illustrate our approach. The figure above focuses on deep learning, specifically on transaction node classification using the ‘isFraud?’ attribute as a label. We’ve structured this as a tripartite type. In contrast, the figure below is tailored for analytics, following a bipartite structure. The main difference lies in where the information is included — either on a node or an edge.

Subtopic 2: Feature Engineering for Transaction Features

Another area of consideration is feature engineering. With the diverse types and channels for sending and receiving money today, we have become more comfortable but also face increased risks. Factors such as device, IP address, card, bank, time, and amount must be carefully considered in feature engineering.

For instance, a graph design focusing on ‘user’-’card’-’merchant’ will have analytics results centered around the middle node, as many users use the card system for transactions. It’s crucial to be mindful of these design choices and interpret analytics results accordingly.

Subtopic 3: Task Setup for Graph Deep Learning

As mentioned earlier, our focus in graph deep learning revolves around the chosen task. Whether it’s node classification for detecting fraud accounts or graph classification to capture the evolving pattern of a client personally, careful task setup is vital.

Your insights on these considerations will contribute significantly to our journey in fraud detection, graph databases, graph analytics, and graph deep learning.

#frauddetection #graphdatabase #graphanalytics #graphdeeplearning #gnn

Thank you for your time and consideration.