Topology Over Predictions: Using Graph Context to Drive Real-Time Lending Decisions
summary
Argues for replacing periodic retraining with continuous graph memory that keeps institutional context by preserving topology rather than flattening it to features. Uses a live Neo4j graph, vector embeddings and vector similarity search plus a policy engine to recalibrate PD and EL as relationships evolve; a pilot shows a merchant score rising from 0.58 by about 12 points and results in 15% fewer false negatives while avoiding retrains and retaining explainability.
AI-written summary · about
chapters (10)
- 0:00Introduction and background
- 1:07Topology vs prediction in finance
- 2:31Institutional amnesia in models
- 3:56Flattening graphs loses meaning
- 5:20Continuous graph reasoning
- 6:03Real-time graph resilience example
- 7:51Reducing false negatives with topology
- 8:36Episodic vs continuous learning
- 9:19Topology for sparse data
- 10:03Explainability and governance
related talks
23mConnecting the Dots With Context Graphs
24mUnlocking AI's Hidden Connections With Graphs
18mConnecting the Dots with Context Graphs
44mWhat Finance Can Learn From Dating Sites
47mWhat Business Can Learn from Dating
01h 00mGraph Mastery to AI
27mEnhancing Financial Analytics With Event-Driven Architectures and Graph Neural Networks
48mOpening Keynote: Exploring Context Graphs: From Data to Decisions - Extended Version