Skip to main content

Async review

Instruction and application
In Progress

Recap core topics:​

  • Unit 3: Identifying and Mitigating Sources of Bias

Unit 3: Identifying and Mitigating Sources of Bias​

In Unit 3, you explored…

  • Sources of bias: How unfairness can enter the ML lifecycle through data collection, preprocessing, model selection, or evaluation choices—and how these factors shape model outcomes for different subgroups.
  • Fairness metrics and evaluation: How to detect algorithmic bias using metrics such as demographic parity, equal opportunity, and predictive parity, and why subgroup-level analysis provides a clearer view of fairness than overall accuracy alone.
  • Explainable AI (XAI) techniques: How tools like SHAP, LIME, and Fairlearn help visualise and interpret model decisions, uncover hidden bias, and strengthen transparency in model evaluation.
  • Bias mitigation strategies: How to apply pre-processing (rebalancing data), in-processing (fairness constraints or regularisation), and post-processing (threshold adjustments) methods to reduce unfair disparities while maintaining strong performance.
  • Fairness documentation and monitoring: How to summarise fairness findings using model audit cards and establish continuous monitoring protocols to track fairness and prevent bias drift as data and demographics evolve.

What is fairness bias in ML?​

  • Fairness bias happens when model outcomes differ unfairly across subgroups.
  • It can emerge at any stage of the ML lifecycle:Data: Underrepresentation or skewed labels.
  • Modelling: Algorithms that favour majority patterns.
  • Evaluation: Metrics that miss subgroup differences.

Why fairness matters​

Biased models can have serious consequences—ranging from ethical risks like discrimination, tocompliance issues that breach regulations, and aloss of trust among users and stakeholders.

Ensuring fairness helps protect both people and organisational credibility.

Detecting bias – looking beyond accuracy​

  • High accuracy doesn’t guarantee fairness.
  • Use fairness metrics such as:Demographic parity – equal outcomes.
  • Equal opportunity – equal true positive rates.
  • Predictive parity – equal precision.

Tools that make bias visible​

Tools like Fairlearn,Aequitas,SHAP, andLIME help uncover how models treat different subgroups. They make fairness measurable and transparent, showing where disparities exist and guiding teams toward more equitable model performance.

Mitigating and monitoring bias​

Once bias is detected, the next step is to reduce its impact while maintaining model performance.

  • Pre-processing: Rebalance or reweight data.
  • **In-processing:**Apply fairness constraints or regularisation.
  • Post-processing: Adjust thresholds or outputs.

Keep fairness accountable​

Use model audit cards to record fairness findings, decisions, and actions. Regularlymonitor model performance as data and user demographics change to ensure fairness remains consistent over time.

Action item: How fair is your model?​

Let’s start with a quick poll to see how you think about fairness and bias in ML systems. These questions explore how you detect, interpret, and address fairness issues in your models.

There are no right or wrong answers—just choose the option that best reflects your approach or experience!