Counterfactual Explanations

1) What it is

• A counterfactual explanation shows how a prediction would change if the input were different.
• It answers:

“What minimal change to this input would flip the model’s decision?”

• Inspired by the idea of counterfactuals in causality: “What would have happened if …?”

---

2) Example

Loan model predicts Denied.
Counterfactual explanation:

• If income increased by $5,000, the prediction would change to Approved.
• If credit score was 650 instead of 600, loan would be approved.

This gives actionable insights — users see what they could change to alter outcomes.

---

3) Formal Definition

Given an instance $x$ with prediction $f(x)$,
a counterfactual is an alternative instance $x’$ such that:

1. $f(x’) \neq f(x)$ (prediction changes).
2. $x’$ is as close as possible to $x$ (minimal change).

Optimization form:

$\min_{x’} \; d(x, x’) \quad \text{s.t.} \quad f(x’) = y_{\text{desired}}x′$

where $d(\cdot)$ = distance metric (e.g., L1, L2, feature-specific cost).

---

4) Properties of Good Counterfactuals

• Validity: Changes prediction successfully.
• Proximity: Minimal change from original.
• Sparsity: Change as few features as possible.
• Actionability: Changes are realistic and controllable (e.g., “get older” is not actionable).
• Diversity: Multiple counterfactuals give users more options.

---

5) Methods to Generate Counterfactuals

• Optimization-based: Solve directly for minimal perturbation.
• Gradient-based search (for differentiable models).
• Generative models (VAE, GAN) to ensure realistic counterfactuals.
• Libraries: alibi, dice-ml.

---

6) Applications

• User-facing transparency:
  • “What do I need to change to get loan approval?”
• Bias/fairness audits:
  • Check if minority groups require unfairly large changes for same decision.
• Model debugging:
  • Understand sensitivity of predictions to features.
• Recourse recommendation:
  • Suggest actionable steps to achieve desired outcomes.

---

7) Limitations

• Can generate unrealistic counterfactuals (e.g., “if gender changed to male, loan approved”).
• Not unique — many possible counterfactuals.
• Requires domain constraints to ensure validity + actionability.

---

Summary

• Counterfactual explanations = “what-if” scenarios showing minimal input changes needed to alter model output.
• Strength: intuitive, actionable for end users.
• Used in finance, healthcare, HR, fairness audits.
• Key challenge: ensure generated counterfactuals are realistic and fair.