Section 2
Taxonomies of Bias
Two primary types of harm
- Allocative harms
- Representational harms
Is this good training data?
✅ Historical Bias
✅ Historical Bias
Normative statement: a misalignment between the world as it is, and your values or objectives
✅ Historical Bias examples
✅ Mitigating Historical Bias
- Can't manage what we can't measure
- Test for model accuracy & allocation across sub-demographics
- Encode 'positive bias' - think affirmative action, for models
- Counterfactual fairness
Can we capture this data well?
✅ Measurement / Sampling Bias
✅ Representation Bias
✅ Measurement Bias
Problems with choosing, collection, or computing features and labels to use in a prediction problem
✅ Measurement Bias
- Measurement process varies across groups
- Quality of data varies across groups
- Defined classification task is an oversimplification
✅ Measurement Bias examples
✅ Mitigating Measurement Bias
- Review sampling methodologies
- Assess proxy quality
- Synthetic data
Can we form useful inferences from our data?
✅ Aggregation Bias
✅ Evaluation Bias
✅ Aggregation Bias
When a one-size-fit-all model is used for groups with different conditional distributions
✅ Aggregation Bias examples
✅ Mitigating Aggregation Bias
- Consider salient differences in sub-demographics for your model
- Add more detailed classifications
- Identify for whom your model performs well
✅ Evaluation Bias
Occurs when the evaluation and/or benchark data for an algorithm doesn’t represent the target population
✅ Evaluation Bias examples
✅ Mitigating Evaluation Bias
- More representative benchmark datasets
- More granular benchmarks (sub-group evaluation)
- Multiple confidence levels
Can we deploy those inferences effectively?
✅ Deployment Bias
✅ Deployment Bias
A mismatch between the problem a model is intended to solve and the way in which it is actually used.
✅ Deployment Bias examples
✅ Mitigating Deployment Bias
- Involve designers, product managers, security people, and other product disciplines in application design
- Conduct UX research on the use of your tool in the real world
- Externalise model uncertainty
- Design failure & recourse flows (first)