Just-So Stories for AI: Explaining Black-Box Predictions
summary
Claims that random forests used inside Stripe Radar are paired with post hoc explanations and a LIME-like pipeline to turn tree predicates into human‑readable rules. The system samples from millions of tree splits (examples cite trees with ~12,000 splits and ~2.5M possible explanations), picks ~1,000 candidates, composes up to ~six‑predicate rules, and ranks by precision for coverage. The result is deployed explanatory output that is persuasive and useful for fraud workflows but acknowledged as plausibly true rather than provably causal.
AI-written summary · about
chapters (29)
- 0:00Introduction
- 0:25Outline
- 2:24Stripe
- 4:02Rules
- 9:03Models
- 12:08Decision Trees
- 15:18Random Forest
- 17:42Explanations
- 20:01Intuition
- 23:41Structure
- 24:15Algorithm
- 26:58Explanation
- 28:56Elephant Trunk
- 30:03Observation
- 30:37Lime
- 33:26AI Rationalisation
- 33:51Frogger
- 34:09Methow submodel interpretability
- 34:22Human interpretability
- 34:44Peter Norvig
- 35:05Roger Sperry
- 37:05Homo Deus
- 37:27Algorithms
- 37:54Explanations are harmful
- 38:48Why explanations are important
- 39:28Human compatible AI
- 39:46Data protection regulation
- 40:43Clarify our ethics
- 41:17Conclusion
related talks
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
33mExplainable AI in the LIME-light
16mBuilding Trust through Explainable AI
32mExplainable AI Explained for Developers
17mChallenges Towards Production-Ready Explainable Machine Learning
29mTowards Explainable Machine Learning
26mRevealing the Unseen - Leveraging XAI for Deeper Data Insights
37mCreating correct and capable classifiers