Revealing the Unseen - Leveraging XAI for Deeper Data Insights
summary
Claims a production XAI pipeline that explains fraud scores for >400M (half‑billion) PayPal accounts, grounding explanations in Shapley values and running daily after pragmatic engineering. Uses TreeSHAP for exact Shapley on tree ensembles, builds a surrogate lightweight model + path‑aggregation trick (grouping identical tree paths via the pigeonhole insight) to cut runtime from infeasible to hours, and validates explanations with a double‑blind analyst experiment yielding statistically significant agreement (p significant) — deployed and working at scale.
AI-written summary · about
chapters (8)
- 0:00Introduction to XAI at PayPal
- 3:58Why explainability matters
- 8:30Shapley values for feature contributions
- 13:12TreeSHAP for efficient Shapley calculation
- 17:31Mapping features to business insights
- 20:03Validating XAI with double-blind test
- 21:08Scaling XAI to production
- 24:27Summary and key takeaways
related talks
18mReal-time Fraud Detection using Deep Learning
35mSHAPtivating Insights: unravelling blackbox AI models
31mBeyond the Black Box: Interpreting ML Models with SHAP
27mInterpretability and explainability in machine learning
30mTrustworthy AI - opening up the black box
39mThe Search for Transparency and Accountability in the Age of...
20mBusiness Value through XAI
33mXAI meets Natural Language Processing