Towards Interpretable Data Science
summary
Argues that explanations must run across the data science cycle and gives concrete systems and numbers: uses Shapley values over provenance compiled into deterministic decomposable circuits via c2d, obtains compilation success ~85% for TPC‑H and ~99% for IMDb, falls back to a linear‑time proxy ranking with precision@5 ≈ 1, and uses a 10s cutoff because exact runs typically finish in ~2.5s. Introduces T‑Rex for repair attribution, a CFDB declarative layer for constrained counterfactuals that enforces integrity constraints (improving retrained models) and reuses generated counterfactuals to speed up analysis.
AI-written summary · about
chapters (9)
- 0:00Interpretable data science
- 1:07Explainable AI for databases
- 8:10Shapley values in queries
- 16:48Visualizing contributions
- 17:09Shapley values in cleaning
- 18:39Explaining EDA notebooks
- 22:11Database techniques for machine learning counterfactuals
- 27:49Counterfactual databases
- 31:00Summary and future
related talks
34mProvenance for Reproducible Data Science
26mDebugging Large-Scale Datalog: A Scalable Provenance Evaluation Strategy
52mData Over Code: A Short Hyperbole
14mDeductive Optimization of Relational Data Storage
17mContext Graphs for Explainable, Decision-Aware AI Agents
17mBring Your Own Data Structures to Datalog
47mDeploying DP: Experiences from the Front Lines and Open Challenges
36mData pipelines engineering made simple with Scala