Balancing scale and interpretability in analytical applications with sklearn and e
summary
PyData Carolinas 2016 describes an implemented stacking framework that partitions a large feature space into analyst-specific subsets, fits scikit-learn pipelines (imputation, normalization, PCA, lasso, random forest, mixed models) as level‑0 regressors, and fits a non-negative lasso combiner on held‑out data. Applied to retail sales (~13 weeks, ~200 stores) for two products, the stacker raises R² above the best base model, yields base‑regressor rankings via likelihood‑ratio tests, and is production‑ready.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 7:29Modeling and collaboration challenges
- 9:18Model stacking overview
- 12:11Building base regressors
- 17:55Combining models into stacker
- 24:46Case studies and results
- 29:28Conclusion
- 31:56Q&A
related talks
44mEnsemble Techniques for High Performance Machine Learning
01h 39mApplying Rigorous Machine Learning Methods in Business Strategy
36mBuild prod ML systems with feature /training /inference pipelines
32mMachine Learning Infrastructure at Stripe: Bridging from Python JVM
57mGood Practices for Applied Machine Learning
33mUsing Scikit-Learn's interface for turning Spaghetti Data Science into Maintainable Software
37mCreating correct and capable classifiers
01h 24mInpsect and try to interpret your scikit-learn machine-learning models