| Adjoint |
Map transferring a linear operation across
a dot product; the transpose for the real Euclidean coordinates
here. |
| Anchor |
Reviewed pair representing the same entity
in two maps for an alignment. |
| Article association |
Recorded candidate link between an article
and a canonical person, with attribution evidence and review state. |
| Attention share |
Allocated scalar story attention divided
by a stated source-period total. |
| Basis |
Independent directions used to express a
space or subspace. |
| Calibration |
Agreement between modeled probabilities
and observed label frequencies for a defined task. |
| Cell |
A specified source-and-decade grouping
used for comparison and support. |
| Centering |
Subtracting the selected reference mean
from observations. |
| Contrast |
A difference from a declared reference; a
person contrast averages supported person-cell minus background
rows. |
| Coordinate |
Amount along a named direction in a
specified coordinate system. |
| Correlation |
Covariance divided by both coordinates’
positive standard deviations. |
| Covariance |
Average product of paired centered
coordinates, with an explicit denominator convention. |
| Crosswalk |
Reviewed table connecting distinct
vocabularies or representations while retaining their provenance. |
| Derivative |
Limiting rate of scalar output change per
input change; a partial derivative varies one input while holding others
fixed. |
| Discrimination |
Ability to distinguish positive and
negative labeled cases, separate from calibration. |
| Eigenvalue and eigenvector |
A stretch factor and direction satisfying
covariance times direction equals factor times direction. |
| Embedding |
Numerical representation of text produced
by a specified model. |
| Frobenius norm |
Square root of the sum of squared matrix
entries. |
| Gradient |
Collection of partial derivatives giving
local changes of an objective with its parameters. |
| Held-out data |
Observations excluded from a stated
fitting or tuning stage and reserved to evaluate it. |
| Incidence table |
Table marking which entities participate
in which articles or events. |
| Intercept |
Baseline linear score when all input
coordinates are zero. |
| Logistic function |
Map from a real score to a value between
zero and one using the exponential function. |
| Mean |
Coordinate sum divided by its observation
count, or a stated weighted counterpart. |
| Norm |
A length; the Euclidean norm here squares
coordinates, adds them, and takes the square root. |
| Normalization |
Dividing a nonzero row by its own length
to retain direction. |
| Ontology |
Reusable vocabulary of concepts and
relationships among their meanings. |
| Orthogonal and orthonormal |
Perpendicular; perpendicular with unit
length. An orthogonal square matrix changes orientation while preserving
lengths. |
| PCA |
Principal component analysis: finding
perpendicular directions of greatest variation in centered
observations. |
| People pattern |
Learned direction of variation across
person profiles. |
| Person profile |
A normalized, background-adjusted
direction of one person’s associated news coverage. |
| Projection |
Measuring and retaining components along
chosen directions. |
| Procrustes alignment |
Orthogonal fit aligning paired anchor rows
by minimizing squared discrepancies. |
| Pseudoinverse |
Generalized inverse giving a least-squares
solution, selecting the shortest when multiple solutions exist. |
| Rank |
Number of independent directions
represented in a matrix. |
| Reconstruction |
Approximation rebuilt from retained
coordinates. |
| Residual |
Original row minus reconstructed row,
before any summary of its size. |
| Ridge penalty |
Added squared-slope cost that trades
prediction fit against coefficient magnitude. |
| Singular value |
Nonnegative scale connecting left and
right directions in an SVD. |
| Slope |
Coefficient multiplying one input
coordinate in a linear prediction. |
| Standard deviation |
Nonnegative square root of variance. |
| Standardization |
Dividing coordinates by their reference
standard deviations; mean subtraction must also be specified. |
| Subspace |
All weighted combinations of selected
directions. |
| Support |
Eligible observed articles satisfying the
stated profile gates; not a guarantee of statistical reliability. |
| SVD |
Singular value decomposition, factoring a
table into left directions, nonnegative scales, and transposed right
directions. |
| Trace |
Sum of a square matrix’s diagonal
entries. |
| Variance |
Average squared centered coordinate value
under a declared denominator convention. |
| Whitening |
Transformation giving unit variances and
zero cross-coordinate covariance in the chosen retained space. |