Counterfactual Explanation Generation

Counterfactual Explanations describe the smallest change to an input that would change the model's prediction — answering "what would need to change for the outcome to be different?" — providing actionable, intuitive explanations that highlight the decision boundary.

Generating Counterfactual Explanations

• Optimization: $\min_{\delta} d(x, x+\delta)$ subject to $f(x+\delta) = y'$ (find the minimum perturbation that changes the prediction).
• Feasibility: Constrain counterfactuals to be realistic/actionable (e.g., can't change age in a loan application).
• Diversity: Generate multiple diverse counterfactuals for richer explanations.
• Methods: DiCE, FACE, Growing Spheres, Algorithmic Recourse.

Why It Matters

• Actionable: Counterfactuals tell users what to change to get a different outcome — directly actionable advice.
• Rights: EU GDPR encourages "right to explanation" — counterfactuals are a natural form of explanation.
• Debugging: In semiconductor AI, counterfactuals reveal which parameters would change a yield prediction.

Counterfactual Explanations are "what would need to change?" — the most actionable form of explanation, showing the minimal path to a different outcome.