Struct FitSet 

pub struct FitSet {
    pub rows: Vec<(Feedback, usize)>,
    pub absent: Vec<Vec<usize>>,
}

A log projected onto one feature order and standardized — exactly what the likelihood sees. Derived at fit time from the log plus a standardizer, and never persisted: the log is the source of truth, this is a view of it.

Fields

rows: Vec<(Feedback, usize)>

Standardized feedback, paired with its session index, in log order.

absent: Vec<Vec<usize>>

Coordinate indices imputed in each row, index-parallel to Self::rows. Empty rows and an empty vector both mean “nothing was imputed”, which is the common case and the one that costs nothing.

A parallel vector rather than a field on the tuple because rows is read positionally in a dozen places (tests, the session layer, the model); widening it would touch all of them to say something only the likelihood needs.

Why the likelihood needs it

An absent coordinate is imputed at the standardizer’s mean, which standardizes to exactly 0 — the honest imputation for “this observation says nothing about that axis”. For a duel that is the end of it: both candidates carry the same absence, so the term cancels in u_a − u_b and the observation is silent about that axis, correctly.

For keep/kill and stars it does not cancel, because there is no second candidate to cancel against. u(x) is compared to a threshold, and a coordinate imputed at zero contributes exactly zero to that sum — so the model reads a patch that might be extreme on the missing axis as though it were average on it, and takes the resulting comparison at full confidence. The information is missing; the certainty should be too, and without this it is not.

Implementations

impl FitSet

pub fn build(log: &ObservationLog, names: &[String], sz: &Standardizer) -> Self

Project every observation onto names and standardize with sz.

Coordinates the observation does not have are imputed at the standardizer’s mean — which standardizes to exactly 0, i.e. “this observation says nothing about that axis”, the honest imputation for a feature that did not exist when the vote was cast. Observations from a legacy standardized log are re-used as-is (they are already z-scores); they are on a different geometry, so the session layer migrates them to raw values first where it can.

pub fn as_is(log: &ObservationLog) -> Self

Take the log’s vectors as already being on the model’s scale (unit tests and synthetic users work directly in standardized space).

pub fn len(&self) -> usize

Number of observations.

pub fn is_empty(&self) -> bool

True when there is nothing to condition on.

pub fn n_sessions(&self) -> usize

Number of distinct sessions referenced (max session index + 1).

Trait Implementations

impl Clone for FitSet

fn clone(&self) -> FitSet

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for FitSet

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for FitSet

fn default() -> FitSet

Returns the “default value” for a type.

impl PartialEq for FitSet

fn eq(&self, other: &FitSet) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for FitSet