OpeningTheBlackBox
summary
Claims that models are no longer uninterpretable, showing a pipeline on a 1.2 million row credit default dataset with 346 features, excluding protected attributes, and using label/one‑hot/target encoding and weight‑of‑evidence. Feature selection uses an ensemble of six importances, permutation tests, and SHAP, leading to a final XGBoost model tuned via H2O AutoML with H2O target encoder. A live streamlit app (≈150 lines) performs what‑if sliders and reproduces a 0.1315 threshold and a 42,000 non‑linear hump in risk.
AI-written summary · about
chapters (15)
- 0:00Model interpretability intro
- 1:30Data and features overview
- 3:21Feature engineering explained
- 4:23Feature selection approaches
- 6:30ML interpretability basics
- 7:56SHAP for model explanation
- 8:24Regulatory compliance
- 9:27Global feature importance
- 10:29Noisy features identification
- 11:35Target encoding
- 13:02SHAP summary plots
- 15:35Dependence plots
- 17:21Streamlit visualization
- 20:07Live Streamlit demo
- 28:11Conclusion & Q&A
related talks
35mSHAPtivating Insights: unravelling blackbox AI models
01h 36mOpen the Black Box: an Introduction to Model Interpretability with LIME and SHAP
30mTraditional & Novel Feature Selection Approaches
31mBeyond the Black Box: Interpreting ML Models with SHAP
47mBreaking Black-Box AI
35mTaking a peek under the hood - Interpreting black box model
39mMachine Learning in Financial Credit Risk Assessment
42mInterpreting Black-Box Models with Applications to Healthcare