Uber uses many methods, including graph learning to improve the quality and relevance of food and restaurant recommendations on its platform. Therefore, when fraudulent behavior, including collusion by cybercriminals, increased, Uber developed a relational graph-learning model to detect colluding users.
Uber applied a Relational Graph Convolutional Networks (RGCN) model on a small sample of data to predict if users are committing fraud. In the user graph, there were two nodes – drivers and riders. Typically, users on a platform are connected via shared information. Hence, Uber found that distinguishing between different connection types helped in amplifying fraud detection signals.
To evaluate the RGCN model performance and utility of the fraud scores, Uber trained the model on four months of historical data up to a specific split date. Subsequently, they tested the model performance with the data collected six weeks after the split date. Further, the fraud scores for users were explicitly generated to compute the precision, recall, and area under the curve. Besides adding two fraud score features to the existing production model, 15% better precision with a minimal increase in false positives was observed.
In the future, Uber plans to identify a more efficient approach for storing massive graphs and conducting distributed training and real-time serving. Furthermore, as driver-rider graphs are densely connected, Uber intends to explore attention-based graph models to make message passing more efficient.
Click here to read the original article published by Uber Engineering
Please give your feedback on this article or share a similar story for publishing by clicking here.
