EXL details GNN fraud detection system with 27-33% detection gains
AI Engineer · youtube · 2026-10-13
EXL's Somya Rai and MD Sharique explain a fraud detection system that adds graph embeddings to an XGBoost classifier, targeting precision within a fixed analyst alert budget. They report 27-33% relative detection improvements over the tabular baseline at matched alert rates.
Key engineering decisions:
- Sampling preserves connectivity: selected users keep full transaction histories so neighborhoods survive
- Leakage controls: monthly graphs, earlier timestamps, training-only preprocessing
- GraphSAGE with mean, max, and std aggregation to preserve unusual neighbors averaging would hide
- Edge attribution linking suspicious scores to specific transactions for analyst review
- Serving: cached representations plus a small online forward pass; moving from transaction-node to user/merchant embeddings cut cache footprint from 60GB to 2-3GB
The talk closes with rolling windows, temporal decay, and millisecond latency tradeoffs rather than a universal benchmark.
More from coding & agent
- Dev spins up a horde of Claudes to hit a one-week deadline — "it worked out" — rickasaurus · 2026-10-13
- Debug Multi-Agent Workflows: Permissions, Handoffs and Channels Beat Model Blame — Al_Grigor · 2026-10-13
- iOS Dev Mocks the AI-Agent Era: HomeView, agents.md, Skills and App Store Rule 2.3bc — Dimillian · 2026-10-13
- Who decides when AI evaluator and human reviewer disagree on an agent call? — AssociationNew7925 · 2026-10-13
- Agentic coding solves the easiest part of engineering and breaks the hardest — cgouguen · 2026-10-13
- Dev builds real-time dictation pipeline: AssemblyAI Sync API plus LLM Gateway rewrite — tech__unicorn · 2026-10-13