What are counterfactual explanations?

Lens adapted to this topic: Limits, assumptions, and practical objections

A critical perspective argues that counterfactuals can look more useful and objective than they are. A suggested feature change may not correspond to a realistic action, may ignore relationships among features or broader causes, and may become misleading when the model changes. Evidence about user value is also limited: one survey reports that only a minority of surveyed methods had been user tested.

• Their usefulness depends on assumptions that feature changes map clearly to real-world actions and that the model is stable, monotonic, and suitably simple.
• Changing model inputs does not necessarily reveal real-world causation; inputs may explain the model’s behavior without being causes of the actual outcome.
• Feature-highlighting explanations can be subjective because they select some relevant features while withholding others.
• A survey of 100 methods reported that only 21% had been user tested, raising questions about how well claimed psychological and practical benefits have been established.

Featured source15  Counterfactual Explanations – Interpretable Machine Learningchristophm.github.io3 sourcesView sources