GoSage: Heterogeneous Graph Neural Network Using Hierarchical Attention for Collusion Fraud Detection
Soumava Ghosh, R. L. Anand, Tanmoy Bhowmik, S. Chandrashekhar · 2023
We propose a graph learning mechanism in a unique but very practical heterogeneous setting, where multiple types of relations can exist between pair of nodes of same or different type. Such problem setting is ubiquitous in digital product/service platform. For example, food delivery or digital payment platform where customer and merchant can be connected by multiple relations such as transactions, shared device and or source/sync of fund etc. Capturing the node representation which encapsulates the multiple facets of information represented by each relation type is vital to detect fraudulent and collusive interaction pattern in the interconnected digital world. We at Gojek, design and deploy a multi-level attention based GNN, which we name as GoSage to intelligently capture the inter-node and inter-relational dependencies in the connected world of our consumers and merchants. We train GoSage in a semi-supervised fashion to detect collusion and fraud in our service and payment network. Through performance comparison as well as ablation study, we establish that our framework is able to outperform state-of-the-art GNN. Multiple experiments and evaluation also establish superior performance even at industrial application scale.