A label is an annotation or target value attached to a data item. It states the outcome a model should learn to predict and, when a reference value becomes available, provides a basis for checking predictions. The first practical question is not simply how to collect labels, but what the product needs them to represent. section 6.2.4 section 14.4.3
One image, several possible labels
For a radiology assistant, doctors might label whether an image shows malignancy, what kind it shows, which region contains it, or exactly which pixels represent it. These choices support different classification, detection, or segmentation tasks. They also demand different tools, expertise, and labeling time. Choose the level of detail from the product goal rather than assuming that more detail is always feasible. section 6.2.4
Before committing to a labeling process, settle the questions that affect the rest of the system:
- What outcome must each data item carry, and where will that value come from? In retail sales forecasting, the item is a date–product–store bundle and the target is units sold; transaction and stock histories are important sources. section 6.7.1 section 6.7
- Can people produce labels at the required detail and scale? Check the available experts, time, tools, instructions, task interface, and quality controls. If precise labels are impractical, consider approaches that can use partial or inaccurate labels. section 6.2.4
- How will label quality be checked? For crowdsourced image-search ranking labels, cross-checks and honeypots can help filter inaccurate labelers; algorithmic validation is another option to consider. section 14.5.2
Labels matter after deployment, too. Once ground truth or feedback arrives, compare it with predictions to calculate model-quality metrics, including regression or classification measures, and inspect performance by segment. If the reference outcome arrives late, monitor data and prediction drift as interim signals rather than treating them as measured prediction quality. section 14.4.3
