Data quality glossary
Key terms from this course, in plain English.
- Active learning
- Choosing the most informative unlabelled examples for humans to label next.
- Cohen's kappa
- A measure of agreement between two labellers that corrects for chance.
- Completeness
- The extent to which required data is present.
- Consistency
- Whether the same thing is recorded the same way across records and systems.
- Data drift
- A shift in the data a model sees compared with its training data.
- Data leakage
- Information unavailable at prediction time influencing training or evaluation.
- Data lineage
- A record of where data came from and how it was transformed.
- Data profiling
- Systematically measuring a dataset's structure, completeness, distributions and anomalies.
- Data-centric AI
- Improving AI by improving data rather than only models.
- Datasheet for datasets
- Documentation of a dataset's motivation, composition, collection, preparation, uses and maintenance.
- Imputation
- Filling missing values using a rule or model.
- Inter-annotator agreement
- How consistently different people label the same items.
- IQR rule
- Flags values more than 1.5 × the interquartile range below the first or above the third quartile.
- Label
- The target answer attached to an example in supervised learning.
- One-hot encoding
- Representing a category as a set of 0/1 columns.
- Outlier
- A value far from most others — which may be an error or a genuine extreme.
- Pipeline
- A chained sequence of preprocessing and modelling steps fitted together.
- Proxy
- A feature that indirectly encodes another attribute, such as a protected characteristic.
- Pseudonymisation
- Replacing identifiers with codes that can only be re-linked using separate information.
- Representativeness
- How well a dataset reflects the population and situations a model will serve.
- Stratified split
- A split that preserves class proportions in each part.
- Target leakage
- Using a feature that is only known after the outcome being predicted.
- Validity
- Whether data follows required formats, types and rules.
- Weak supervision
- Labelling data at scale with rules or heuristics that are noisy but cheap.