AIF-C01 Practice Question 9: Unrepresentative Training Data Bias
AWS Certified AI Practitioner (AIF-C01) · Test 1 · question 9 of 10
After launch, a health-insurance risk model is found to be noticeably less accurate for applicants under 25 and over 70 than for everyone else. An investigation shows that most of the training records came from middle-aged customers. What is the most appropriate way to address this fairness issue?
Show answer and explanation
Answer:
This is bias that comes from the data. A model learns mostly from the patterns it sees most often, so a group with little training data tends to get worse predictions. The fix starts at the source: make the training data representative of the people the model will serve. Then measure performance for each group, not just overall, to confirm the gap has closed (option C). Fairness, meaning the impact on different groups of stakeholders, is one of the core dimensions of responsible AI.
- A is wrong because a high overall accuracy can hide poor accuracy for small groups. An average doesn't measure fairness. - B is wrong because removing a sensitive attribute doesn't guarantee fairness. Other features can act as proxies for age, and the under-representation is still in the data. - D is wrong because a bigger model trained on the same skewed data still has little to learn from for those groups, and it's harder to explain.
Reference: Pre-training Data Bias