Fraud Detection: Using Relational Graph Learning to Detect Collusion

Uber logo
Uber
•
Xinyu Hu, Chengliang Yang, Ankur Sarda, Ankit Jain, Piero Molino
•11 min read•advanced•
--
•View Original
ApacheApache SparkConvolutional Neural NetworksDeep LearningDGLGraph Neural NetworksNeural Networks

Overview

The article discusses the application of relational graph learning, specifically relational graph convolutional networks (RGCN), to detect collusion in fraudulent activities within the Uber platform. It outlines a case study demonstrating how graph learning can enhance fraud detection by leveraging user connections and shared information.

What You'll Learn

1

How to apply relational graph convolutional networks for fraud detection

2

Why distinguishing connection types improves fraud detection accuracy

3

How to construct a user-relationship graph for fraud analysis

Prerequisites & Requirements

  • Understanding of graph learning concepts
  • Familiarity with the DGL library(optional)

Key Questions Answered

How does relational graph learning improve fraud detection?
Relational graph learning enhances fraud detection by utilizing user connections and shared information to identify patterns of collusion. By analyzing the relationships between users, the model can detect clusters of fraudulent activity, leading to more accurate predictions of fraud.
What is the role of edge types in RGCN for fraud detection?
Edge types in RGCN provide relation-specific transformations that enhance the model's ability to differentiate between various user interactions. This allows the model to weigh messages from connected nodes differently, improving the accuracy of fraud detection by considering the context of each connection.
What are the performance metrics used to evaluate the RGCN model?
The RGCN model's performance is evaluated using precision, recall, and AUC metrics. During testing, it was observed that adding two fraud score features improved precision by 15% with minimal increase in false positives, demonstrating the model's effectiveness in identifying fraudulent users.

Key Statistics & Figures

Precision improvement
15%
This improvement was observed when two fraud score features were added to the existing production model.
Feature ranking
4th and 39th
The two fraud scores learned from RGCN were ranked 4th and 39th among 200 features in the downstream model.

Technologies & Tools

Library
Dgl
Used for constructing and training the relational graph convolutional networks.

Key Actionable Insights

1
Implementing relational graph convolutional networks can significantly enhance fraud detection capabilities.
By leveraging user connections and shared information, organizations can identify fraudulent patterns more effectively, leading to reduced financial losses.
2
Utilizing edge types in graph models allows for more nuanced fraud detection.
This approach helps differentiate between various user interactions, which can improve the model's predictive accuracy and reduce false positives.
3
Regularly updating the user-relationship graph is crucial for maintaining detection accuracy.
As user behavior changes, the graph must reflect these updates to ensure that the fraud detection model remains effective.

Common Pitfalls

1
Overlooking the importance of edge types in graph models can lead to suboptimal fraud detection.
Without considering the context of user interactions, the model may fail to accurately identify fraudulent behavior, resulting in higher false positive rates.

Related Concepts

Graph Learning Techniques
Fraud Detection Methodologies
Relational Graph Convolutional Networks