Applying Graph Computing for Risk Control in Wing Pay: Architecture, Algorithms, and Future Directions

The talk begins with an overview of Wing Pay, a subsidiary of China Telecom that operates payment, e‑commerce, and credit services, and outlines the major fraud risks in each scenario, including account theft, money laundering, subsidy abuse, and loan fraud.

It then introduces the graph‑based risk control framework, explaining that entities such as accounts, devices, bank cards, and IDs become nodes, while relationships like transactions, logins, and red‑packet transfers become edges, forming a large‑scale graph stored in a distributed graph database.

The speaker describes two families of graph algorithms: traditional graph‑theoretic methods (e.g., connectivity, Louvain, label propagation, PageRank) and deep learning‑based graph neural network (GNN) models such as Node2Vec, GCN, GAT, and GraphSAGE.

Three algorithmic pipelines are presented: SubGraph‑based subgraph mining, MetaPath pattern matching, and GNN‑based learning. Real‑time subgraph partitioning is achieved by caching group information in Redis, allowing millisecond‑level risk interception.

Case studies illustrate how large‑scale fraud rings and long‑chain collusion can be uncovered by visualizing graph structures that are invisible to rule‑based systems, highlighting the importance of homophily, scarcity, and timeliness in fraud detection.

The future outlook covers needs for massive distributed native graph databases with real‑time ingestion, automated rule mining from expert knowledge, explainable alerts, distributed GNN training frameworks (e.g., DGL, PyG), multimodal heterogeneous data fusion, and graph‑based federated learning to preserve privacy.

Q&A discusses the integration of GNN embeddings with GBDT models and the scale of entities and edges used in anti‑money‑laundering graphs, which is reduced to a few million after filtering normal users.

RESULT SET = SET A MATCHED ∪ SET B MATCHED ∪ … ∪ SET N MATCHED