Machine Learning (ML) · Advanced level

Advanced level · Module 2: Interpretability, Fairness and Documentation · Lesson 3 of 4

Mitigating Unfairness

Reduce unfair outcomes through data, modelling and decision-threshold choices, and check the costs.

In this lesson you will

  • Apply data, in-training and post-processing mitigations
  • Evaluate the trade-offs of each mitigation

Key concepts

data collection and reweightingconstrained traininggroup-specific thresholdstrade-off evaluation

The lesson

Mitigation can happen at several stages: collect more representative data or reweight it; train with fairness constraints; or adjust decision thresholds after training.

Every mitigation has costs — accuracy, complexity or legal questions about treating groups differently. Measure them and decide with stakeholders rather than optimising one number.

Sometimes the right decision is not to deploy a model for a purpose, or to keep a person making the final decision with the model as one input.

Step by step

  1. Reweight training data and retrain.
  2. Try a threshold adjustment.
  3. Compare fairness and accuracy for each option.
  4. Write a recommendation.

Practice activity

Recommend a mitigation approach for the credit model, with evidence.

Expected result: A recommendation comparing options on fairness and performance, including the option of not automating.

Check your understanding

  1. Name three stages where unfairness can be mitigated.

    Show answer

    Data, training and post-processing (thresholds).

  2. Is deploying always the right outcome?

    Show answer

    No; sometimes a model should not be used or should only support human decisions.

Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.