Edge 263: Local Model-Agnostic Interpretability Methods: Counterfactual Explanations
Counterfactual explanations as an ML interpretability method, Google's StylEx and Microsoft's DiCE implementation
In this issue:
An overview of local-interpretability methods based on counterfactual explanations.
Google’s StylEx method for generation visual explanations in image classifiers
Microsoft’s open source implementation of the DiCE method.
💡 ML Concept of the Day: : Local Model-Agnostic Interpretability Methods: Counterfactual Explanations
In the last few editions of this newsletter, we have been discussing local model-agnostic interpretability methods which focus on deriving explanations based on the outcome of a single prediction. One of the most interesting methods in this area of ML interpretability is known as counterfactual explanations. From causality theory, we know counterfactual explanations as a predicate that describes the relationship between two variables in the following form: “If X had not occurred, Y would not have occurred. How different would the world be if the Soviet Union would have landed in the moon first. Who knows but makes for an interesting counterfactual analysis. Consider a more pragmatic example of a home loan application rejected by an ML algorithm. A counterfactual analysis can reveal than an income increase is the most likely factor to increase the chances to getting a loan next time.
