Data quality is how well data serve their intended use. For machine learning, more observations do not necessarily mean better data: useful training data must be accurate, relevant, and broad enough to represent operating conditions. section 11.6.3 section 13.5.2
Assess training data along several dimensions:
- Accuracy and consistency: incorrect labels and inconsistent measurements can weaken learned representations. section 11.6.3
- Relevance and coverage: data should reflect the conditions and population where a model will operate; more examples from a narrow setting do not fill coverage gaps. section 11.6.3
- Informational value: duplication, errors, and narrow distributions can add volume without adding much new information. section 13.5.2
For example, a pedestrian detector trained mostly on clear daytime images may fail at night or in severe weather. More daytime images do not remedy that gap. Validation on data from the same source as training may also miss distribution shifts, so production performance needs monitoring as conditions change. section 11.6.3 section 11.6.3














