Glossary and notation
Core vocabulary¶
- Condition
- A predicate on one descriptor attribute, such as
age > 40. - Description / intension
- The logical expression defining membership. In the standard searches it is a nonempty conjunction of supplied conditions on distinct features.
- Cover / extension
- The set of rows matching a description. Two descriptions may have the same cover while receiving different generalization-aware scores.
- Subgroup
- Depending on context, the description, its cover, or both. This book states which aspect matters when the distinction affects the analysis.
- Coverage / relative size
- The proportion of dataset rows matching a description. An encoder’s
Coverageobject instead denotes a representation of condition masks. - Target
- The outcome, distribution, or model behavior under investigation. In the API, a fitted object storing positional per-row information and reference statistics.
- Descriptor
- An attribute permitted to define subgroup membership, as distinct from model or target attributes.
- Quality function
- The definition used to rank candidates, always maximized by the search. Its numerical scale depends on the measure and target units.
- Support
- Often the number/fraction of matching rows in frequent pattern mining. In
binary result tables,
supportis the fraction both covered and positive. - Target share
- The positive-label fraction within a binary subgroup.
- Lift
- Subgroup target share divided by dataset target share. The API’s
LiftQFinstead scores the share difference, which is order-equivalent when the reference positive share is fixed and nonzero. - Refinement / specialization
- A description made more specific by adding conditions. Its cover cannot grow.
- Generalization
- A less specific description. For generalization-aware measures here, it is obtained only by removing conditions; interval widening is not included.
- Empty description
- A conjunction with no conditions, covering the full dataset. It is used as the reference generalization but is not returned by standard discovery searches.
- Optimistic estimate
- An upper bound on the quality attainable by a refinement, allowing a branch to be pruned without visiting every descendant. It is not an estimate of the subgroup’s expected population effect.
- Anti-monotone constraint
- Once it fails, it also fails for every refinement. Minimum covered count is an example; minimum target share is not.
- Top-k
- Up to the highest-scoring eligible descriptions. The number returned is not the number evaluated.
- Beam
- A bounded set of descriptions retained for further refinement by a heuristic search. A discarded branch can contain the optimum.
- Complement
- Dataset rows outside the subgroup. Unlike the dataset reference, it does not contain the subgroup itself.
- Diversity
- A property of a selected set, defined through descriptions, covered rows, attribute use, or compression rather than only individual quality.
- Generalization awareness
- Scoring a description’s improvement over simpler descriptions. It does not imply low overlap with other selected descriptions or independent validation.
- Exceptional Model Mining (EMM)
- Discovery where interestingness concerns a fitted model or relationship among target attributes, rather than only a single target’s level.
- Valuation basis
- An aggregatable summary of rows from which a model or quality can be computed, such as count, sum, and cross-product statistics used by GP-growth.
- Holdout validation
- Testing fixed discovery candidates on data not used to choose them.
- Randomization null model
- The distribution of the statistic over randomized row sets constrained as specified, such as equal size and optionally equal class/stratum counts.
- Adjusted p-value
- A p-value transformed for a specified multiple-testing procedure over the tested family. It is not the probability that the subgroup is false.
Notation used in the book¶
| Symbol | Meaning |
|---|---|
| , $N= | D |
| , $n= | G |
| Subgroup complement | |
| , | Positive counts in dataset and subgroup |
| , | Positive shares |
| , | Numeric target means |
| , | Class-share vector and numeric mean vector |
| , | Quality and optimistic estimate |
| Size exponent in a standard quality function | |
| , | Final result limit and larger candidate pool for validation |
| , | Random draws and number scoring at least as high as observed |
| Significance level, or size weight in the SubROC formula when explicitly stated | |
| Class-balance exponent in SubROC |
The same letters can have different local uses in publications. Constructor keyword arguments are preferable when a paper’s notation and parameter order can be confused.