Practical Tips for Interpreting Machine Learning Models - Patrick Hall, H2O.ai
- Observability
- H2O Driverless AI
Overview
In this conference talk, H2O.ai's Patrick Hall condenses machine learning interpretability into nine practical tips. He covers sensitivity analysis and random data attacks for stress-testing models, why explanation-generating software itself needs to be tested for stability and accuracy, and the importance of combining global techniques, like variable importance, with local ones, like SHAP, rather than relying on just one. He also walks through monotonically constrained XGBoost paired with SHAP explanations, decision tree surrogate models paired with partial dependence and ICE plots, and practical caveats for using LIME.
Key takeaways
- Why you should run sensitivity analysis, testing your model on values outside its training range, and random data attacks to see how it behaves on data it wasn't built for
- Why explanation-generating software should be tested for stability and accuracy, not assumed to be correct out of the box
- Why you need both global interpretability techniques (like variable importance or partial dependence) and local ones (like SHAP) to understand a model fully
- How combining monotonically constrained XGBoost with SHAP explanations gives you a free, deployable, and explainable model
- How a decision tree surrogate model paired with partial dependence and ICE plots can reveal important variable interactions
- What LIME is good for, where it can fail, especially in highly nonlinear regions, and tricks like discretizing inputs or averaging multiple local models to make it more reliable
Start owning your AI