MAIE5103 Chap.6 Auditing and Mitigating Bias
Auditing and Mitigating Bias
Document before you measure
The audit half of the fairness lecture starts with documents rather than metrics. A datasheet records why a dataset exists, what it contains, how it was collected and preprocessed, what it may be used for and how it is maintained.
The data audit then asks about provenance, gaps and intended use: who gathered the data and under what licence, which groups, languages and devices are missing and how often, and what the data must never be used for.
A model card records intended use, metrics by group, evaluation data and caveats, and the lecture's phrase for it is a structured confession of scope rather than a certificate.
Numbers with sample sizes and intervals
Every rate on a fairness dashboard is an estimate.
The lecture's rule is that each one carries its sample size, its confidence interval and its date, that the interval goes on the difference between groups rather than on each group separately, and that an interval covering zero is reported as inconclusive rather than fair.
Slicing by intersections finds the worst-served cell but shrinks cells until intervals explode, so small cells are suppressed or widened under a documented minimum size.
Repairs at three stages, and a human
Pre-processing reweights or repairs data but cannot fix a biased label. In-processing builds a fairness penalty or an adversary into training and optimises the metric you wrote rather than the harm.
Post-processing keeps the model and changes thresholds, which is transparent but needs the attribute at decision time. Human review can correct a model or rubber-stamp it. Monitoring and a five-step incident response close the loop, with ownership placed in operations.
What this chapter covers
- 01
Datasheets, data audits and model cards
- 02
Provenance, gaps and intended use
- 03
Single-attribute, intersectional and clustered slices
- 04
Intervals on the difference and the small-cell rule
- 05
Fairness dashboards with named owners
- 06
Pre-processing, in-processing and post-processing
- 07
Threshold choice and human-in-the-loop design
- 08
Limits of debiasing, monitoring and incident response
Worked example · free
Rewrite a dashboard tile so it can be trusted
- 1Replace the single score with named metrics by group.
- 1Add slices, sample sizes and intervals.
- 1Add drift signals and say what an alert should do.
- 1Name the owner and the response path.
Key terms
- Data Audit
- A review of a dataset's provenance, gaps and intended use before any model is trained on it.
- Reweighting
- A pre-processing repair that gives under-represented combinations of inputs, labels and groups more weight in training.
- Adversarial Debiasing
- An in-processing method that trains the predictor to defeat a second model trying to recover the protected attribute from its outputs.
- Small-Cell Rule
- A reporting rule that suppresses or widens results for subgroups below a documented minimum size so that noise is not mistaken for a finding.
- Fairness Dashboard
- A monitoring display of named metrics by group, with slices, sample sizes, intervals, drift signals and a named owner.
- Fairness Incident
- A harm event involving a group disparity, handled through detection, containment, documentation, remediation and notification.
Auditing and Mitigating Bias FAQ
What is the difference between pre-processing, in-processing and post-processing?
They act at different stages. Pre-processing changes the training data, by reweighting or repairing features, and cannot fix a biased label. In-processing changes training itself, adding a fairness penalty or an adversary, and optimises only the metric you specify.
Post-processing keeps the trained model and changes the decision threshold for each group, which is transparent but needs the protected attribute when decisions are made.
How do I know whether a fairness gap is real or just noise?
Put a confidence interval on the difference between the groups' rates, using a binomial or bootstrap method, and report the sample sizes beside every rate. If the interval covers zero, the honest word is inconclusive. Testing many slices inflates false gaps, so decide the primary attribute in advance, and when a cell is tiny either collect more data or widen the group.
Does adding a human reviewer make an automated decision fair?
Not by itself. Reviewers can correct a model or defer to it under time pressure, which is automation bias. The lecture asks what the reviewer sees, when review is mandatory, how overrides are recorded and who is reviewed more often. Debiasing methods have similar limits: the lecture calls them controls rather than certificates, because a model can find another correlate.
Assessment move
Compute one confidence interval on a difference by hand, using the normal approximation, until the steps feel routine; then use it whenever a case gives you two group rates and a small sample. For the mitigation families, make a three-row table from memory with the stage, what each method needs and what it cannot fix, and check it against this chapter. Those two artefacts cover most of what written work on audits asks.
Working through Auditing and Mitigating Bias in MAIE5103? Sia is AskSia’s AI Arts and Humanities tutor — ask any MAIE5103 Auditing and Mitigating Bias question and get a clear, step-by-step explanation grounded in how MAIE5103 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.