Global methods describe the average behavior of a machine learning model. The counterpart to global methods are local methods. Global methods are often expressed as expected values based on the distribution of the data. For example, the partial dependence plot, a feature effect plot, is the expected prediction when all other features are marginalized out. Since global interpretation methods describe average behavior, they are particularly useful when the modeler wants to understand the general mechanisms in the data or debug a model.
In this book, you will learn about the following model-agnostic global interpretation techniques:
- The partial dependence plot is a feature effect method.
- Accumulated local effect plots is another feature effect method that works when features are dependent.
- Feature interaction (H-statistic) quantifies to what extent the prediction is the result of joint effects of the features.
- Functional decomposition is a central idea of interpretability and a technique that decomposes the complex prediction function into smaller parts.
- Permutation feature importance measures the importance of a feature as an increase in loss when the feature is permuted.
- Global surrogate models replaces the original model with a simpler model for interpretation.
- Prototypes and criticisms are representative data point of a distribution and can be used to enhance interpretability.