A model built from historical approvals has low recall for one group.
Investigate and improve the cause.
Choose two appropriate responses.
Inspect group counts, labels, missingness, and selection, then compare resampling, weighting, and label changes on performance and fairness.
Detailed explanation
The harm is hidden and information is lost.
Data-generation and measurement bias can be found.
Bias and proxy effects may remain.
Performance and fairness effects are both measured.
Check official documentation
Compare the following points with the official documentation.
AWS公式AIF-C01 Domain 4.1のバイアス、データ品質、公平性対策を確認する。
Official references
What to confirm
偏りの原因をデータ生成過程まで追い、緩和策の副作用も評価できる。
Key points
- Label quality
- Resampling
- Proxy
Notes
- Environment: AWS公式AIF-C01試験ガイドとAWS公式ドキュメントの確認
- Compare the definition with the AIF-C01 objective and the service documentation before applying it.
- Keep the concept separate from a particular answer choice or implementation step.
Foundation review
Sampling bias
A difference between the observed training sample and the population the system will serve.
Resampling
Changing which records are selected or repeated so training groups receive different weights.