Towards Interpretable Data Science
summary
Argues that explanations must run through the whole data‑science cycle and demonstrates concrete tools: uses Shapley values with provenance and knowledge compilation to compute tuple attributions, achieving compilation success for ~85% of TPC‑H outputs and ~99% for IMDB, with a hybrid cutoff near 2.5s and precision@5 ≈1; visualizes results via ShapGraph; applies attribution to repairs in T-REX (pruned sampling makes it practical); and speeds constrained counterfactual analysis and reuse with CFDB/CFQL, improving CERTIFAI-style fairness checks.
AI-written summary · about
chapters (8)
- 0:00Introduction to explainable AI
- 2:58Feature attribution and Shapley
- 7:20Database provenance for XAI
- 13:20Scalable explanation techniques
- 18:19Explaining data preparation
- 20:05Explaining data exploration
- 23:55Constraint-based counterfactuals
- 29:37Declarative counterfactual system
related talks
27mInterpretability and explainability in machine learning
26mRevealing the Unseen - Leveraging XAI for Deeper Data Insights
39mExplainable AI as a Conventional Data Analysis Tool
34mInterpretable AI or How I Learned to Stop Worrying and Trust AI
16mBuilding Trust through Explainable AI
31mBeyond the Black Box: Interpreting ML Models with SHAP
37mCreating correct and capable classifiers
28mThe Need for Explainable AI