Minimising bias in data collection and preprocessing
Instruction and application
CompleteLineage and metadata reduce the chance that opaque, unfair data reaches your model. Bias can still enter during collection and preprocessing—sometimes silently until real users are harmed.
What is bias?
Bias is systematic skew in data, labels, features or outcomes. It may come from unrepresentative samples, inconsistent measurement, proxy variables or human labelling choices. The damage often appearslate, after deployment.
Common bias types (with mitigations)
Data bias
Training data does not reflect the real world. Mitigation: broaden sources, languages and geographies; document exclusions; monitor slice metrics.
Sampling bias
The sample differs from the target population. Mitigation: design sampling plans for coverage; weight or resample carefully; document known gaps.
Confirmation bias
Teams select or interpret data to match prior beliefs. Mitigation: pre-register evaluation slices; independent review; adversarial testing.
Measurement bias
Collection or labelling rules differ across groups or time. Mitigation: standardise instruments and label guides; audit labelers; harmonise historical schema changes.
Algorithmic bias
Models amplify inequality even when raw inputs look neutral. Mitigation: fairness constraints, post-processing, regular audits, diverse review panels.
Human bias
Annotators or product owners encode stereotypes into labels or objectives. Mitigation: diverse labeling teams, rubrics, inter-rater reliability, appeals channels.
Key points
Ignoring bias risks unfair decisions, reputational crises and regulatory enforcement. Mitigation is technical and organisational: metrics alone cannot replace accountable process design.
Mitigation strategies in depth
Diverse data collection
Scenario: A loan model trained mainly on historical approvals inherits past discrimination.Response: add regulated alternative data (where permitted), community finance signals and carefully validated features; pair with fairness testing.
Key point
“More data” is not automatically better—more representative, well-documented data is.
Data augmentation for fairness
Scenario: An image classifier confuses occupational roles with gender because training photos skew stereotypical.Response: oversample underrepresented roles, generate synthetic-but-realistic images where policy allows and re-balance evaluation sets.
Bias detection tooling
Scenario: A résumé screener recommends men for leadership roles disproportionately.Response: run toolkits such as Fairlearn or AI Fairness 360 to measure disparity; apply constraints, reweighting or remove proxy features; revalidate after each retrain.
Ethical foundations beyond math
- Fairness and equity: test outcomes and errors across protected and operational slices where law and ethics apply.
- Transparency and explainability: pair preprocessing documentation with model explanations for denials, care pathways or risk scores.
- Accountability: assign owners for dataset approvals, publish audit cadence and create escalation when metrics breach thresholds.
Moderation and language models illustrate the stakes: automated systems can disproportionately flag dialects or communities when training data and policies are not continuously reviewed.
Action item: Quiz
- A. The model samples mini-batches randomly each epoch
- B. The training data does not represent the population the model will serve
- C. The dataset is too large to fit in memory
- D. The optimiser converged to a local minimum
Feedback: Coverage of people, places and time periods is a design choice—get it wrong and metrics lie.
- A. Drop all rows with missing values without analysis
- B. Train only on the majority class for speed
- C. Carefully apply augmentation, weighting or resampling with documented evaluation
- D. Remove fairness metrics to improve accuracy
Feedback: Augmentation and weighting can help but need governance—synthetic data and duplicates carry their own risks.