This chapter covers
- Characteristics of deep neural networks
- How to implement deep neural networks that are inherently black-box models
- Perturbation-based model-agnostic methods that are local in scope, such as LIME, SHAP and anchors
- How to interpret deep neural networks using LIME, SHAP, and anchors
- Strengths and weaknesses of LIME, SHAP, and anchors
In the previous chapter, we looked at tree ensembles, especially random forest models, and learned how to interpret them using model-agnostic methods that are global in scope, such as partial dependence plots (PDPs) and feature interaction plots. We saw that PDPs are a great way of understanding how individual feature values impact the final model prediction at a global scale. We were also able to see how features interact with each other using the feature interaction plots and how they can be used to expose potential issues such as bias. PDPs are easy and intuitive to understand, but their major drawback is that they assume features are independent of each other. In addition, higher-order feature interactions cannot be visualized using feature interaction plots.