Understanding Decisions Driven by Big Data
summary
Argues that a simple evidence counterfactual—computing the minimal set of features whose removal flips a model’s decision—lets teams see why massive, sparse models decide as they do. Applied to a KDD corporate‑residence fraud detector it pinpoints invoices (e.g. Fairway Golf Club) that drive suspicious labels and exposes false positives; applied to PNAS Facebook‑likes results it lists likes (e.g. Glee, Lady Gaga) that push a sexual‑orientation or IQ prediction. Enables a practical cloak likes privacy control.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:42The Dirty Secret
- 1:26Finegrained Data
- 2:36Reasons for Understanding Decisions
- 4:19The Explanation Trick
- 5:55Corporate Residence Fraud Detection
- 7:31Evidence Counterfactual
- 10:13Transparency
- 11:13Evidence Counter Factual
related talks
31mHow to Overcome the Apathy of Big Data
10mIs Bigger Really Better? Predictive Analytics with Fine-grained Behavior Data
43mKeynote: Consequences of an Insightful Algorithm
25mClaudia Perlich, Chief Scientist, Dstillery, NYC 2017
47mAI Ethics, Strategic Decisioning and Game Theory
01h 12mObscured by algorithms: how to prove "yes" if computer says "no"
01h 16mData protection and Big Data: what about regulating Big Data analytics?
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python