2.3. Data Integrity and Modeling Readiness
💡 First Principle: A model is only as fair and reliable as the data it learns from. "Garbage in, garbage out" is literally true in ML—but the more insidious problem is bias in, bias amplified out. Data integrity isn't just a quality concern; it's an ethical, legal, and business-critical concern. The exam tests your ability to detect and mitigate these issues using AWS tools.
Think of data integrity like a building inspection before construction. You can build a beautiful skyscraper (model), but if the foundation (data) has cracks — missing values, label errors, demographic imbalances — the structure will fail under load. AWS provides tools at each inspection stage: Glue DataBrew for profiling, SageMaker Data Wrangler for transformation, and SageMaker Clarify for bias detection. The exam expects you to know which tool fits which inspection stage.
What happens when a hiring model trains on historical data where 85% of hired candidates were male? The model learns that gender is predictive of hiring success—not because it's true, but because the data reflects historical bias. Without bias detection and mitigation, you've automated discrimination. This is why SageMaker Clarify exists, and it's why Domain 1 explicitly tests pre-training bias detection.
Consider data integrity as a three-layer shield: the outer layer checks quality (is the data correct and complete?), the middle layer checks fairness (does the data represent all groups equitably?), and the inner layer checks compliance (does the data handling meet regulatory requirements?). All three must pass before training begins.
⚠️ Common Misconception: Clarify's pre-training bias detection and post-training bias detection are not interchangeable. Pre-training analysis (DPL, CI metrics) runs on the dataset before any model exists — it catches data collection problems. Post-training analysis (DPPL, DI metrics) evaluates the model's predictions against protected groups. The exam asks you to select the right phase: if no model has been trained yet, post-training metrics don't apply.
