View a markdown version of this page

RAIDP03-BP02 Minimize unwanted bias in your datasets - Responsible AI Lens
Implementation considerationsResources

RAIDP03-BP02 Minimize unwanted bias in your datasets

When assessing the quality of a dataset, determine whether it appropriately represents the demographics of the expected range of system users. Consider datasets that include self-reported demographic labels. Calculate if datasets contain sufficient representation across demographic groups to enable statistically valid fairness assessments or fairness outcomes.

Level of risk exposed if this best practice is not established: High

Implementation considerations

  1. Analyze the demographic composition of your datasets to identify which groups may be over- or under-represented for your use case.

  2. Consider using self-reported demographic labels. For example, consider using survey responses or user-provided information rather than algorithmic or human predictions of demographic information.

  3. Calculate statistical power for each demographic group in your evaluation datasets by working backwards from your release criteria. For instance, determine whether you have enough examples per group to answer each release criteria question with the required statistical confidence.

  4. Address representation gaps by collecting additional data from underrepresented groups or using techniques like stratified sampling, where a population is divided into subgroups, or "strata," based on shared characteristics, and then a random sample is taken from each subgroup to verify representation.

  5. Validate that your bias mitigation efforts don't introduce new fairness concerns. For example, check if balancing one demographic dimension inadvertently creates imbalances across intersectional groups.

Resources

Related documents: