Interpretable versus black-box machine learning

Interpretable and simple models such as linear regression make it easy to assess the possibility of improving them, finding issues with them such as biases that need to be detected and removed, and building trust in using such models. However, to achieve higher performance, we usually don’t stop with these simple models and rely on complex or so-called black-box models. In this section, we will review some of the interpretable models and then introduce techniques you can use to explain your black-box models.

Interpretable machine learning models

Linear models such as linear and logistic regression, shallow decision trees, and Naive Bayes classifiers are examples of simple and interpretable methods (Figure 6.1). We can easily extract the contribution of features in predictions of outputs for these models and identify opportunities for improving their performance, such as by adding or removing features or changing feature normalization...