What are data labels?
Data labels are the known answers attached to the examples a model learns from. A label can be a category (“approved” or “denied”), a number (a sale price), a box around an object in a photo, or a tag on a span of text. In supervised learning, the model sees the input and the label together, and training adjusts it until its predictions match the labels as often as possible.
Labels come from records the business already keeps (an invoice was paid late or on time), from people who review examples by hand (the work of data annotation), or from a trusted rule that serves as ground truth. The set of possible labels in a classification task is the set of classes.
Why data labels matter
A model can only be as good as its labels. If a collections agency labels call recordings by outcome (“promise to pay”, “dispute”, “wrong number”) and its labelers disagreed on a quarter of the calls, the model learns that disagreement. Poor label quality is a more common cause of a disappointing first model than the choice of algorithm.
What to watch for
- Write a labeling guide with examples before anyone starts, and measure how often two labelers agree.
- Confirm each label was recorded before the event it predicts, or the model learns from information it will not have in production.
- Count the examples per class, because a class with only a handful of labels is hard to learn.