In the context of Explainable AI, and counterfactual reasoning in particular, generating a counterfactual example can be made by using two types of strategies:
- Case-based, where some repository of data points such as a dataset is used i order to find the counterfactual example, given a certain query point. The advantage of this method is that it can be very effective if there is a rich and relevant dataset available, the obvious disadvantage is the fact that the quality of the counterfactual depends on the quality and completeness of the dataset used.
- Instance-based, where the query point’s features are modified in order to generate a counterfactual, usually using optimization or search techniques to achieve a good counterfactual. The advantage is the flexibility introduced by the fact that it doesn’t need any additional data; the main disadvantage is that the counterfactual may be less interpretable, and requires careful tuning to ensure that the generated counterfactuals are good (meaning plausible and actionable).
tags: counterfactual resources: Guidotti - Counterfactual explanations and how to find them - literature review and benchmarking