Contrastive Explanation

Contrastive Explanations explain a model's prediction by contrasting it with an alternative outcome — answering "why outcome A instead of outcome B?" by identifying features that are present for A (pertinent positives) and absent features that would lead to B (pertinent negatives).

Components of Contrastive Explanations

• Foil: The alternative outcome to contrast against (e.g., "why class A and not class B?").
• Pertinent Positives (PP): Minimal features present in the input that justify the predicted class.
• Pertinent Negatives (PN): Minimal features absent from the input whose presence would change the prediction.
• CEM: Contrastive Explanation Method finds both PPs and PNs using optimization.

Why It Matters

• Human-Like: Humans naturally explain by contrast — "I chose A over B because of X."
• Focused: Contrastive explanations highlight only the discriminating features, not all features.
• Diagnostic: For manufacturing, "why did this wafer fail instead of pass?" is a natural contrastive question.

Contrastive Explanations are "why this and not that?" — focusing explanations on the differences that discriminate between the predicted and alternative outcomes.