Construction and Application of a Financial Event Knowledge Graph

Background : In Ant Financial’s business loop, users, merchants, institutions, and their flows, products, and funds form a closed operational cycle that is influenced by external macro‑economic, policy, and institutional events. To improve perception of these events and operational efficiency, a financial event knowledge graph was built for wealth‑management and risk scenarios.

Key Challenges : (1) High business complexity increases knowledge‑modeling cost; (2) High coverage and precision required for information extraction; (3) Graph density must be sufficient to support event‑logic operations, avoiding sparse redundant nodes.

Construction Pipeline :

Knowledge Modeling – Identify entities, attributes, and relationships to define entity types and meta‑concepts, with events defined as multi‑entity combinations.

Event and Logic Understanding – Extract event instances from data sources, link them to meta‑concept constraints, abstract event concepts, and integrate them into the graph.

Graph‑Enhanced Search & Content Generation – Use events and their linked concepts as inputs for recommendation algorithms and generate short texts for analysts or end‑users.

Graph Overview : The graph consists of three layers – instance, concept, and meta‑concept – with dense connections among entities (companies, products, industries) and events (macro, industry, and enterprise‑level).

Applications :

Financial Institution Risk : Real‑time T+0 risk perception and T‑N risk prediction using the graph to detect high‑risk company events, generate alerts, and support downstream risk models.

Wealth‑Management Search & Recommendation : Treat events as high‑information‑density content for recommendation, requiring multi‑element event extraction, entity linking, and audience segmentation.

Industry‑Chain Analysis : Link upstream and downstream products to identify micro‑enterprise impacts under macro events, using taxonomy expansion to build a product knowledge base.

Technical Insights : Challenges addressed include complex label taxonomy with contrastive loss, handling long‑tail events via event‑graph sharing, and reducing annotation noise through continual learning.

Summary : The graph contains seven major event categories, 46 secondary concepts, and over 700 effective meta‑concepts, enabling fine‑grained constraints and richer business automation.

Q&A Highlights : Sample sizes (≈120k company events, ≈30k market events), single‑graph storage of instances, concepts, and meta‑concepts, and the three‑layer architecture supporting both inference and user‑facing services.