Buy Book Buy

Chapter 5 Interpretable Models

The easiest way to achieve interpretability is to use only a subset of algorithms that create interpretable models. Linear regression, logistic regression and the decision tree are commonly used interpretable models.

In the following chapters we will talk about these models. Not in detail, only the basics, because there is already a ton of books, videos, tutorials, papers and more material available. We will focus on how to interpret the models. The book discusses linear regression, logistic regression, other linear regression extensions, decision trees, decision rules and the RuleFit algorithm in more detail. It also lists other interpretable models.

All interpretable models explained in this book are interpretable on a modular level, with the exception of the k-nearest neighbors method. The following table gives an overview of the interpretable model types and their properties. A model is linear if the association between features and target is modelled linearly. A model with monotonicity constraints ensures that the relationship between a feature and the target outcome always goes in the same direction over the entire range of the feature: An increase in the feature value either always leads to an increase or always to a decrease in the target outcome. Monotonicity is useful for the interpretation of a model because it makes it easier to understand a relationship. Some models can automatically include interactions between features to predict the target outcome. You can include interactions in any type of model by manually creating interaction features. Interactions can improve predictive performance, but too many or too complex interactions can hurt interpretability. Some models handle only regression, some only classification, and still others both.

From this table, you can select a suitable interpretable model for your task, either regression (regr) or classification (class):

Algorithm Linear Monotone Interaction Task
Linear regression Yes Yes No regr
Logistic regression No Yes No class
Decision trees No Some Yes class,regr
RuleFit Yes No Yes class,regr
Naive Bayes No Yes No class
k-nearest neighbors No No No class,regr

You could argue that both logistic regression and Naive Bayes allow linear explanations. However, this is only true for the logarithm of the target: Increasing a feature by one point increases the logarithm of the target probability by a certain amount assuming all other features remain the same.