12 Interpretability and explainability

 

Hilo is a (fictional) startup company trying to shake up the online second home mortgage market. A type of second mortgage known as a home equity line of credit (HELOC) allows customers to borrow intermittently using their house as collateral. Hilo is creating several unique propositions to differentiate itself from other companies in the space. The first is that it integrates the different functions involved in executing a second mortgage, including a credit check of the borrower and an appraisal of the value of the home, in one system. Second, its use of machine learning throughout these human decision-making processes is coupled with a maniacal focus on robustness to distribution shift, fairness, and adversarial robustness. Third, it has promised to be scrutable to anyone who would like to examine the machine learning models it will use and to provide avenues for recourse if the machine’s decisions are problematic in any respect. Imagine that you are on the data science team assembled by Hilo and have been tasked with addressing the third proposition by making the machine learning models interpretable and explainable. The platform’s launch date is only a few months away, so you had better get cracking.

12.1   The different types of explanations

12.1.1    Personas of the consumers of explanations

12.1.2    Dichotomies of explanation methods

12.1.3    Conclusion

12.2   Disentangled representation

12.3   Explanations for regulators

12.3.1    k-Nearest neighbor classifier

12.3.2    Decision trees and Boolean rule sets

12.3.3    Logistic regression

12.3.4    Generalized additive models

12.3.5    Generalized linear rule models

12.3.6    Deletion diagnostics and influence functions

12.4   Explanations for decision makers

12.4.1    Global model approximation

12.4.2    LIME

12.4.3    Partial dependence plots

12.4.4    SHAP

12.4.5    Saliency maps

12.4.6    Prototypes

12.5   Explanations for affected users

12.6   Quantifying interpretability

12.7   Summary