Course contentsShow
← Machine Learning and Deep Learning
Lesson 3314 of 3,53872. Bias and FairnessFree lesson

Reject Option Classification

Withhold predictions in a boundary region and manually review or apply group-aware rules for fairness.

Reject Option Classification

What you'll learn: How to defer uncertain predictions near the decision boundary to improve fairness by withholding automated decisions and applying manual or group-aware review.

The Core Idea

Reject Option Classification is a post-processing technique that tackles fairness by creating a "gray zone" around your model's decision boundary. Instead of letting the model make borderline calls that might be biased, you withhold predictions for cases where the model is uncertain—then either manually review them or apply group-aware fairness rules.

Think of it like a referee calling for instant replay: when a play is too close to call confidently, you don't guess—you review carefully or apply a tie-breaking rule that ensures fair treatment across teams.

How It Works

  1. Identify the uncertainty region: Define a threshold range around your decision boundary (e.g., if probability is between 0.4 and 0.6 for a binary classifier).

  2. Reject predictions in that zone: For samples falling in this region, the model says "I'm not sure enough to decide automatically."

  3. Apply fairness-aware resolution:

    • Manual review: A human expert makes the final call
    • Group-aware rules: Use protected group membership to flip predictions that disadvantage underrepresented groups while keeping others unchanged

This approach targets the region where most biased errors occur—near the boundary where the model lacks confidence.

When to Use It

Reject Option Classification shines when:

  • You can afford human review for a subset of cases
  • Errors near the boundary disproportionately harm specific groups
  • You've already trained your model and need a fairness fix without retraining

Key Takeaway: Reject Option Classification improves fairness by withholding uncertain predictions near the decision boundary and applying manual review or group-specific rules, targeting the region where biased errors most often occur.