Ground truth (machine learning)
Reality. The thing that actually happened. For example, consider a binary classification model that predicts whether a student in their first year of university will graduate within six years. Ground truth for this model is whether or not that student actually graduated within six years.1
We assess model quality against ground truth. However, ground truth is not always completely, well, truthful. For example, consider the following examples of potential imperfections in ground truth:1
- In the graduation example, are we certain that the graduation records for each student are always correct? Is the university’s record-keeping flawless?
- Suppose the label is a floating-point value measured by instruments (for example, barometers). How can we be sure that each instrument is calibrated identically or that each reading was taken under the same circumstances?
- If the label is a matter of human opinion, how can we be sure that each human rater is evaluating events in the same way? To improve consistency, expert human raters sometimes intervene.