Explainable AI for Practitioners

Chapter 3. Explainability for Tabular Data

Much of the success of deep learning has focused on unstructured data like images, text, audio, and video; however, the vast majority of machine learning models in production are built with tabular data in mind. Think of all the data contained in relational databases and spreadsheets composed of numeric and categorical feature sets. These are examples of structured data and make up the vast majority of real-world AI use cases. In this chapter, we’ll examine explainability techniques that are most often used when working with tabular data, like Shapley values, permutation feature importance, tree interpreters, and various versions of partial dependence plots.

Permutation Feature Importance

Here’s what you need to know about permutation feature importance:

• Once a model has been fit to the training data, the permutation importance for a single feature measures the decrease in a model score when that feature value is randomly shuffled.

• By shuffling the values of a given feature, you destroy the model’s ability to make meaningful predictions using that feature. If the model predictions suffer and the model score is much worse, then the information provided by that feature must have been important to the model when making predictions. On the other hand, if the change in the model score is negligible then that feature isn’t as important.

Pros	Cons
It’s easy to implement. Scikit-learn provides a nice, easy-to-use library for ...