On the Art and Science of Machine Learning Explanations

Multiple explanation methods exist, but each has different strengths: The paper examines five major approaches (decision trees, ICE plots, LIME, partial dependence plots, and Shapley explanations) that "vary in terms of scope, fidelity, and suitable application domain," meaning organizations need to choose the right tool for their specific situation rather than relying on a one-size-fits-all approach.

Some explanatory methods have stronger theoretical foundations than others: The paper distinguishes between methods that are "accepted tools of the trade" versus those "rigorously derived and backed by long-standing theory," suggesting that practitioners should understand the scientific credibility behind the explanation method they're using to assess AI decisions.

Traditional error measurements alone are insufficient for responsible AI: The paper emphasizes that explanatory methods go "beyond the error measurements and plots traditionally used," indicating that knowing a model performs well statistically doesn't tell stakeholders why it makes specific decisions—critical information for identifying bias or unfair outcomes.

Practical guidance and reproducibility matter for implementation: The paper provides "real-world usage recommendations supported by a use case and public, in-depth software examples," recognizing that responsible AI requires not just theoretical knowledge but actionable guidance that organizations can actually implement and verify.