Application of Knowledge Graphs in Risk Control at Wing Payment

The talk, titled “Knowledge Graph in Wing Payment Risk Control,” was delivered by Director Xu Dehua of the Risk Management Department of Wing Payment, a China Telecom subsidiary, and covered three main topics: the risk‑control knowledge graph, its practical applications, and future outlook.

Risk control is defined as the management of financial risks in three core business scenarios—payment, e‑commerce, and credit—each presenting distinct threats such as money‑laundering, account theft, fraudulent merchants, and intermediary fraud.

The presentation described the sophisticated black‑market (黑产) ecosystem that supplies resources, scripts, and services for large‑scale fraud, highlighting its massive scale (over a trillion yuan loss and more than 1.6 million workers).

Wing Payment’s risk‑control system integrates a comprehensive knowledge graph as a core component, collecting multimodal data (behavioral logs, identity images, text forms, transaction records, and third‑party data) and feeding it into AI engines, rule‑based modules, and a full‑life‑cycle management platform.

Specific AI techniques include visual anti‑fraud computer‑vision models for identity verification, unsupervised methods for unlabeled e‑commerce data, supervised learning for known risk patterns, and federated learning for cross‑company collaboration, with the knowledge graph serving as the central data backbone.

The knowledge graph has been built across three dimensions—payment, finance, and telecom anti‑fraud—reaching billions of nodes and edges, and differs from generic graphs by employing a top‑down, strong‑schema construction, high data quality, and millisecond‑level real‑time decision requirements.

Application layers provide visual exploration tools for business users, a rich algorithm library (traditional graph algorithms such as Louvain, WCC, label propagation, and deep‑learning methods like GCN, GAT, GraphSAGE), and support for subgraph mining, meta‑path pattern matching, and graph neural networks, illustrated through case studies on money‑laundering rings, intermediary fraud, and multi‑modal fraud detection.

Future directions focus on integrating large language models (e.g., GPT) for knowledge representation, enhancing real‑time data ingestion, fusing multimodal data (images, audio, video) into the graph, and advancing federated graph learning while addressing security, privacy, and fairness concerns.

The Q&A revealed that the system serves tens of millions of users, handles peak loads of around 10,000 QPS, and achieves accuracy rates above 99% for critical fraud‑detection scenarios.

The presentation concluded with a brief promotional segment highlighting Wing Payment’s end‑to‑end fraud‑prevention solutions, including OCR‑based identity verification and comprehensive risk‑control platforms.