Responsible Preparation for Datasets
Balanced data
The story: A shoe company designs shoes by measuring feet, but they only measured adult men. Their shoes fit adult men well and hurt everyone else. The higher the stakes (say, medical shoes), the worse that mistake is.
In AI/AWS terms: Responsible models need balanced datasets that represent every relevant group. Balance matters most in high-stakes areas like hiring, lending, and criminal justice. Use SageMaker Clarify to find imbalance and SageMaker Data Wrangler to fix it.
For the exam: Clarify finds data imbalance. Data Wrangler fixes it.
Step 1: Inclusive and diverse collection
The story: The fix starts with who you measure. Go to schools, offices, sports clubs, and care homes, in different cities. And it's not only about people: a shoe for hiking needs data from mud, rock, and snow, not only city pavement.
In AI/AWS terms: Collect from a diverse range of sources, viewpoints, and demographics. A model trained mostly on middle-aged people will be less accurate for younger and older people. Diversity matters for any topic, not only data about people.
For the exam: Diverse collection comes first. A group missing from the data will be served poorly by the model.
Step 2: Data curation
The story: Once you have the measurements, you clean up mistakes and put everything in the same units. If you only managed to measure a few children, you create realistic extra child sizes based on the real ones. And every season, you check the numbers are still balanced.
In AI/AWS terms: Data curation is labeling, organizing, and preprocessing:
- Preprocessing: cleaning, normalization, and feature selection to remove bias and errors.
- Augmentation: generate new examples of underrepresented groups when there isn't enough real data.
- Regular auditing: keep checking that the data stays balanced and fair.
For the exam: Not enough data for a group → augmentation, which generates new examples for underrepresented groups.
Balance for the intended use case
The story: If you're designing shoes for toddlers, you want toddler feet in your data. Adding lots of adult feet "for balance" would make things worse, not better.
In AI/AWS terms: Balance depends on the purpose. A system about cancer in children should be built on data about children, not adults.
For the exam: "Balanced" means representative of the intended use case, not evenly split across everyone.