Explaining Models built in H2O-3
- Observability
- H2O-3
Overview
In this webinar, an H2O.ai principal data scientist walks through model explainability in H2O-3, the company's open-source distributed machine learning platform. After covering the core ideas behind responsible AI, she trains a baseline GBM model on the Lending Club loan dataset and uses the h2o.explain interface to generate global and local explanations, including SHAP summary plots, partial dependence plots, and individual conditional expectation (ICE) plots. It's a practical, code-along session using a shared Colab notebook for understanding why a model makes the predictions it makes.
Key takeaways
- How H2O-3's h2o.explain function generates global explanations (for the whole model) or local explanations (for individual rows) with a single line of code
- How to train a baseline gradient boosting model on the Lending Club loan dataset and evaluate it with a confusion matrix, learning curve, and AUC on train, validation, and test sets
- How to read a variable importance plot and a SHAP summary plot, including what the color and horizontal position of each point represent
- How partial dependence plots show a feature's average marginal effect on predictions, and why you should check the underlying data histogram before trusting the tail of the curve
- How ICE plots reveal individual-level variation, and interactions, that a partial dependence plot's single average line can hide
- How to pull single-function outputs like the variable importance or SHAP plots on their own, and generate local explanations for a specific row
Start owning your AI