Concept reference
On this page
Use these definitions to check an explanation, then return to the worked example. A term is understood when you can apply it to an unfamiliar observation, not merely repeat the definition.
| Term | Meaning in this book | Experiment |
|---|---|---|
| Observation unit | The entity or event represented by one analytical row | Questions and scope |
| Variable | A recorded attribute whose meaning, type and unit must be defined | Data types |
| Key | A field or combination identifying an observation under a stated uniqueness rule | Validation |
| Provenance | Traceable information about source, selection, acquisition and transformation | Acquisition |
| Schema | A contract for fields, types and relevant constraints | Validation |
| Array shape | The length of every array dimension | Array structure |
| Axis | A numbered dimension whose observational meaning you must state | Numerical computation |
| Broadcasting | Combining compatible array shapes without explicitly repeating every input | Array structure |
| View | An array representation that can share its source’s underlying data | Array structure |
| Reduction | A calculation that combines values along one or more axes | Numerical computation |
| Rank | Number of independent directions represented, estimated numerically for a matrix | Matrices |
| Residual | Observed value minus fitted value, or a specified equation’s discrepancy | Mathematical reasoning |
| Population | The set of units the question concerns | Sampling |
| Sample | The observed subset, obtained through a particular selection process | Sampling |
| Conditional probability | Probability under the restriction that specified information or an event is given | Probability |
| Standard deviation | A measure of spread among values, in their original unit | Statistics |
| Standard error | Variation of an estimator under a stated sampling repetition or model | Uncertainty |
| Label alignment | Matching labelled observations before combining their values | Pandas selection |
| Missing value | An unresolved recorded value, whose reason may require additional metadata | Cleaning |
| Cardinality | The intended one-to-one, one-to-many or other matching relationship between keys | Joins |
| Long table | A representation in which a variable such as year appears as values rather than repeated field names | Reshaping |
| Denominator | The reference population or exposure relative to which a rate or proportion is defined | Comparisons |
| Confounding | A data-generating structure in which an observed association can mix relationships relevant to the causal question | Comparisons |
| Resampling | Repeatedly drawing from observations under a stated scheme to investigate an estimator | Statistics |
| Resampling in time | Regrouping observations into defined time intervals; distinct from statistical resampling | Time |
| Baseline | A simple specified procedure used as an evaluation reference | ML questions |
| Feature | An input measurement or representation available under the task’s information rules | Data handoff |
| Target | The quantity or label the prediction task seeks to estimate | Data handoff |
| Leakage | Information entering learning or evaluation that would be unavailable under the intended use or independence boundary | Data handoff |
| Token | A text unit defined by a particular representation procedure | Text records |
| Document frequency | The number of documents containing a specified token, counted once per document | Text records |
| Tensor | In the computational examples, a multidimensional array with explicitly named sample, feature or other dimensions | Data handoff |